Estimate fan duty from observed RPM maxima

- NvidiaPowerBenchResult gains ServerPower *BenchmarkServerPower
- RunNvidiaPowerBench samples IPMI idle before Phase 1 and loaded via
  background goroutine throughout Phase 2 ramp
- renderPowerBenchReport: new "Server vs GPU Power Comparison" table
  with ratio annotation (✓ match / ⚠ minor / ✗ over-report)
- renderPowerBenchSummary: server_idle_w, server_loaded_w, server_delta_w,
  server_reporting_ratio keys

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.gitignore

@@ -2,3 +2,4 @@
 .DS_Store
 dist/
 iso/out/
+build-cache/

audit/go.mod

@@ -5,22 +5,18 @@ go 1.25.0
 replace reanimator/chart => ../internal/chart
 
 require (
-	github.com/go-analyze/charts v0.5.26
+	modernc.org/sqlite v1.48.0
 	reanimator/chart v0.0.0-00010101000000-000000000000
 )
 
 require (
 	github.com/dustin/go-humanize v1.0.1 // indirect
-	github.com/go-analyze/bulk v0.1.3 // indirect
-	github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 // indirect
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
-	golang.org/x/image v0.24.0 // indirect
 	golang.org/x/sys v0.42.0 // indirect
-	modernc.org/libc v1.70.0 // indirect
+	modernc.org/libc v1.72.0 // indirect
 	modernc.org/mathutil v1.7.1 // indirect
 	modernc.org/memory v1.11.0 // indirect
-	modernc.org/sqlite v1.48.0 // indirect
 )

audit/go.sum

@@ -1,37 +1,51 @@
-github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
-github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
-github.com/go-analyze/bulk v0.1.3 h1:pzRdBqzHDAT9PyROt0SlWE0YqPtdmTcEpIJY0C3vF0c=
-github.com/go-analyze/bulk v0.1.3/go.mod h1:afon/KtFJYnekIyN20H/+XUvcLFjE8sKR1CfpqfClgM=
-github.com/go-analyze/charts v0.5.26 h1:rSwZikLQuFX6cJzwI8OAgaWZneG1kDYxD857ms00ZxY=
-github.com/go-analyze/charts v0.5.26/go.mod h1:s1YvQhjiSwtLx1f2dOKfiV9x2TT49nVSL6v2rlRpTbY=
-github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 h1:DACJavvAHhabrF08vX0COfcOBJRhZ8lUbR+ZWIs0Y5g=
-github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0/go.mod h1:E/TSTwGwJL78qG/PmXZO1EjYhfJinVAhrmmHX6Z8B9k=
+github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17kjQEVQ1XRhq2/JR1M3sGqeJoxs=
+github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
+github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
 github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
-github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
-github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=
 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=
-github.com/stretchr/testify v1.11.1 h1:7s2iGBzp5EwR7/aIZr8ao5+dra3wiQyKjjFuvgVKu7U=
-github.com/stretchr/testify v1.11.1/go.mod h1:wZwfW3scLgRK+23gO65QZefKpKQRnfz6sD981Nm4B6U=
-golang.org/x/image v0.24.0 h1:AN7zRgVsbvmTfNyqIbbOraYL8mSwcKncEj8ofjgzcMQ=
-golang.org/x/image v0.24.0/go.mod h1:4b/ITuLfqYq1hqZcjofwctIhi7sZh2WaCjvsBNjjya8=
+golang.org/x/mod v0.33.0 h1:tHFzIWbBifEmbwtGz65eaWyGiGZatSrT9prnU8DbVL8=
+golang.org/x/mod v0.33.0/go.mod h1:swjeQEj+6r7fODbD2cqrnje9PnziFuw4bmLbBZFrQ5w=
+golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
+golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.42.0 h1:omrd2nAlyT5ESRdCLYdm3+fMfNFE/+Rf4bDIQImRJeo=
 golang.org/x/sys v0.42.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
-gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
-gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
-modernc.org/libc v1.70.0 h1:U58NawXqXbgpZ/dcdS9kMshu08aiA6b7gusEusqzNkw=
-modernc.org/libc v1.70.0/go.mod h1:OVmxFGP1CI/Z4L3E0Q3Mf1PDE0BucwMkcXjjLntvHJo=
+golang.org/x/tools v0.42.0 h1:uNgphsn75Tdz5Ji2q36v/nsFSfR/9BRFvqhGBaJGd5k=
+golang.org/x/tools v0.42.0/go.mod h1:Ma6lCIwGZvHK6XtgbswSoWroEkhugApmsXyrUmBhfr0=
+modernc.org/cc/v4 v4.27.3 h1:uNCgn37E5U09mTv1XgskEVUJ8ADKpmFMPxzGJ0TSo+U=
+modernc.org/cc/v4 v4.27.3/go.mod h1:3YjcbCqhoTTHPycJDRl2WZKKFj0nwcOIPBfEZK0Hdk8=
+modernc.org/ccgo/v4 v4.32.4 h1:L5OB8rpEX4ZsXEQwGozRfJyJSFHbbNVOoQ59DU9/KuU=
+modernc.org/ccgo/v4 v4.32.4/go.mod h1:lY7f+fiTDHfcv6YlRgSkxYfhs+UvOEEzj49jAn2TOx0=
+modernc.org/fileutil v1.4.0 h1:j6ZzNTftVS054gi281TyLjHPp6CPHr2KCxEXjEbD6SM=
+modernc.org/fileutil v1.4.0/go.mod h1:EqdKFDxiByqxLk8ozOxObDSfcVOv/54xDs/DUHdvCUU=
+modernc.org/gc/v2 v2.6.5 h1:nyqdV8q46KvTpZlsw66kWqwXRHdjIlJOhG6kxiV/9xI=
+modernc.org/gc/v2 v2.6.5/go.mod h1:YgIahr1ypgfe7chRuJi2gD7DBQiKSLMPgBQe9oIiito=
+modernc.org/gc/v3 v3.1.2 h1:ZtDCnhonXSZexk/AYsegNRV1lJGgaNZJuKjJSWKyEqo=
+modernc.org/gc/v3 v3.1.2/go.mod h1:HFK/6AGESC7Ex+EZJhJ2Gni6cTaYpSMmU/cT9RmlfYY=
+modernc.org/goabi0 v0.2.0 h1:HvEowk7LxcPd0eq6mVOAEMai46V+i7Jrj13t4AzuNks=
+modernc.org/goabi0 v0.2.0/go.mod h1:CEFRnnJhKvWT1c1JTI3Avm+tgOWbkOu5oPA8eH8LnMI=
+modernc.org/libc v1.72.0 h1:IEu559v9a0XWjw0DPoVKtXpO2qt5NVLAnFaBbjq+n8c=
+modernc.org/libc v1.72.0/go.mod h1:tTU8DL8A+XLVkEY3x5E/tO7s2Q/q42EtnNWda/L5QhQ=
 modernc.org/mathutil v1.7.1 h1:GCZVGXdaN8gTqB1Mf/usp1Y/hSqgI2vAGGP4jZMCxOU=
 modernc.org/mathutil v1.7.1/go.mod h1:4p5IwJITfppl0G4sUEDtCr4DthTaT47/N3aT6MhfgJg=
 modernc.org/memory v1.11.0 h1:o4QC8aMQzmcwCK3t3Ux/ZHmwFPzE6hf2Y5LbkRs+hbI=
 modernc.org/memory v1.11.0/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=
+modernc.org/opt v0.1.4 h1:2kNGMRiUjrp4LcaPuLY2PzUfqM/w9N23quVwhKt5Qm8=
+modernc.org/opt v0.1.4/go.mod h1:03fq9lsNfvkYSfxrfUhZCWPk1lm4cq4N+Bh//bEtgns=
+modernc.org/sortutil v1.2.1 h1:+xyoGf15mM3NMlPDnFqrteY07klSFxLElE2PVuWIJ7w=
+modernc.org/sortutil v1.2.1/go.mod h1:7ZI3a3REbai7gzCLcotuw9AC4VZVpYMjDzETGsSMqJE=
 modernc.org/sqlite v1.48.0 h1:ElZyLop3Q2mHYk5IFPPXADejZrlHu7APbpB0sF78bq4=
 modernc.org/sqlite v1.48.0/go.mod h1:hWjRO6Tj/5Ik8ieqxQybiEOUXy0NJFNp2tpvVpKlvig=
+modernc.org/strutil v1.2.1 h1:UneZBkQA+DX2Rp35KcM69cSsNES9ly8mQWD71HKlOA0=
+modernc.org/strutil v1.2.1/go.mod h1:EHkiggD70koQxjVdSBM3JKM7k6L0FbGE5eymy9i3B9A=
+modernc.org/token v1.1.0 h1:Xl7Ap9dKaEs5kLoOQeQmPWevfnk/DM5qcLcYlA8ys6Y=
+modernc.org/token v1.1.0/go.mod h1:UGzOrNV1mAFSEB63lOFHIpNRUVMvYTc6yu1SMY/XTDM=

audit/internal/app/app.go

@@ -30,7 +30,9 @@ var (
 	DefaultRuntimeLogPath   = DefaultExportDir + "/runtime-health.log"
 	DefaultTechDumpDir      = DefaultExportDir + "/techdump"
 	DefaultSATBaseDir       = DefaultExportDir + "/bee-sat"
-	DefaultBenchmarkBaseDir = DefaultExportDir + "/bee-benchmark"
+	DefaultBeeBenchBaseDir  = DefaultExportDir + "/bee-bench"
+	DefaultBeeBenchPerfDir  = DefaultBeeBenchBaseDir + "/perf"
+	DefaultBeeBenchPowerDir = DefaultBeeBenchBaseDir + "/power"
 )
 
 type App struct {
@@ -84,6 +86,7 @@ type installer interface {
 	InstallToDisk(ctx context.Context, device string, logFile string) error
 	IsLiveMediaInRAM() bool
 	LiveBootSource() platform.LiveBootSource
+	LiveMediaRAMState() platform.LiveMediaRAMState
 	RunInstallToRAM(ctx context.Context, logFunc func(string)) error
 }
 
@@ -108,6 +111,10 @@ func (a *App) LiveBootSource() platform.LiveBootSource {
 	return a.installer.LiveBootSource()
 }
 
+func (a *App) LiveMediaRAMState() platform.LiveMediaRAMState {
+	return a.installer.LiveMediaRAMState()
+}
+
 func (a *App) RunInstallToRAM(ctx context.Context, logFunc func(string)) error {
 	return a.installer.RunInstallToRAM(ctx, logFunc)
 }
@@ -117,7 +124,8 @@ type satRunner interface {
 	RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaBenchmark(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error)
-	RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
+	RunNvidiaPowerBench(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error)
+	RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error)
 	RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
 	RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
@@ -139,7 +147,6 @@ type satRunner interface {
 	RunFanStressTest(ctx context.Context, baseDir string, opts platform.FanStressOptions) (string, error)
 	RunPlatformStress(ctx context.Context, baseDir string, opts platform.PlatformStressOptions, logFunc func(string)) (string, error)
 	RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
-	RunHPL(ctx context.Context, baseDir string, opts platform.HPLOptions, logFunc func(string)) (string, *platform.HPLResult, error)
 }
 
 type runtimeChecker interface {
@@ -191,6 +198,7 @@ func (a *App) RunAudit(runtimeMode runtimeenv.Mode, output string) (string, erro
 	}
 	result := collector.Run(runtimeMode)
 	applyLatestSATStatuses(&result.Hardware, DefaultSATBaseDir, a.StatusDB)
+	writePSUStatusesToDB(a.StatusDB, result.Hardware.PowerSupplies)
 	if health, err := ReadRuntimeHealth(DefaultRuntimeJSONPath); err == nil {
 		result.Runtime = &health
 	}
@@ -562,16 +570,23 @@ func (a *App) RunNvidiaBenchmark(baseDir string, opts platform.NvidiaBenchmarkOp
 
 func (a *App) RunNvidiaBenchmarkCtx(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
-		baseDir = DefaultBenchmarkBaseDir
+		baseDir = DefaultBeeBenchPerfDir
 	}
 	return a.sat.RunNvidiaBenchmark(ctx, baseDir, opts, logFunc)
 }
 
-func (a *App) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+func (a *App) RunNvidiaPowerBenchCtx(ctx context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultBeeBenchPowerDir
+	}
+	return a.sat.RunNvidiaPowerBench(ctx, baseDir, opts, logFunc)
+}
+
+func (a *App) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
 	if strings.TrimSpace(baseDir) == "" {
 		baseDir = DefaultSATBaseDir
 	}
-	return a.sat.RunNvidiaOfficialComputePack(ctx, baseDir, durationSec, gpuIndices, logFunc)
+	return a.sat.RunNvidiaOfficialComputePack(ctx, baseDir, durationSec, gpuIndices, staggerSec, logFunc)
 }
 
 func (a *App) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -738,13 +753,6 @@ func (a *App) RunNCCLTestsResult(ctx context.Context) (ActionResult, error) {
 	return ActionResult{Title: "NCCL bandwidth test", Body: body}, err
 }
 
-func (a *App) RunHPL(ctx context.Context, baseDir string, opts platform.HPLOptions, logFunc func(string)) (string, *platform.HPLResult, error) {
-	if a == nil {
-		return "", nil, fmt.Errorf("app not configured")
-	}
-	return a.sat.RunHPL(ctx, baseDir, opts, logFunc)
-}
-
 func (a *App) RunFanStressTestResult(ctx context.Context, opts platform.FanStressOptions) (ActionResult, error) {
 	path, err := a.RunFanStressTest(ctx, "", opts)
 	body := formatFanStressResult(path)
@@ -934,6 +942,41 @@ func bodyOr(body, fallback string) string {
 	return body
 }
 
+// writePSUStatusesToDB records PSU statuses collected during audit into the
+// component-status DB so they are visible in the Hardware Summary card.
+// PSU status is sourced from IPMI (ipmitool fru + sdr) during audit.
+func writePSUStatusesToDB(db *ComponentStatusDB, psus []schema.HardwarePowerSupply) {
+	if db == nil || len(psus) == 0 {
+		return
+	}
+	const source = "audit:ipmi"
+	worstStatus := "OK"
+	for _, psu := range psus {
+		if psu.Status == nil {
+			continue
+		}
+		slot := "?"
+		if psu.Slot != nil {
+			slot = *psu.Slot
+		}
+		st := *psu.Status
+		detail := ""
+		if psu.ErrorDescription != nil {
+			detail = *psu.ErrorDescription
+		}
+		db.Record("psu:"+slot, source, st, detail)
+		switch st {
+		case "Critical":
+			worstStatus = "Critical"
+		case "Warning":
+			if worstStatus != "Critical" {
+				worstStatus = "Warning"
+			}
+		}
+	}
+	db.Record("psu:all", source, worstStatus, "")
+}
+
 func ReadRuntimeHealth(path string) (schema.RuntimeHealth, error) {
 	raw, err := os.ReadFile(path)
 	if err != nil {

audit/internal/app/app_test.go

@@ -122,6 +122,7 @@ func (f fakeTools) CheckTools(names []string) []platform.ToolStatus {
 type fakeSAT struct {
 	runNvidiaFn               func(string) (string, error)
 	runNvidiaBenchmarkFn      func(string, platform.NvidiaBenchmarkOptions) (string, error)
+	runNvidiaPowerBenchFn     func(string, platform.NvidiaBenchmarkOptions) (string, error)
 	runNvidiaStressFn         func(string, platform.NvidiaStressOptions) (string, error)
 	runNvidiaComputeFn        func(string, int, []int) (string, error)
 	runNvidiaPowerFn          func(string, int, []int) (string, error)
@@ -154,6 +155,13 @@ func (f fakeSAT) RunNvidiaBenchmark(_ context.Context, baseDir string, opts plat
 	return f.runNvidiaFn(baseDir)
 }
 
+func (f fakeSAT) RunNvidiaPowerBench(_ context.Context, baseDir string, opts platform.NvidiaBenchmarkOptions, _ func(string)) (string, error) {
+	if f.runNvidiaPowerBenchFn != nil {
+		return f.runNvidiaPowerBenchFn(baseDir, opts)
+	}
+	return f.runNvidiaFn(baseDir)
+}
+
 func (f fakeSAT) RunNvidiaTargetedStressValidatePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
 	if f.runNvidiaTargetedStressFn != nil {
 		return f.runNvidiaTargetedStressFn(baseDir, durationSec, gpuIndices)
@@ -161,7 +169,7 @@ func (f fakeSAT) RunNvidiaTargetedStressValidatePack(_ context.Context, baseDir
 	return f.runNvidiaFn(baseDir)
 }
 
-func (f fakeSAT) RunNvidiaOfficialComputePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ func(string)) (string, error) {
+func (f fakeSAT) RunNvidiaOfficialComputePack(_ context.Context, baseDir string, durationSec int, gpuIndices []int, _ int, _ func(string)) (string, error) {
 	if f.runNvidiaComputeFn != nil {
 		return f.runNvidiaComputeFn(baseDir, durationSec, gpuIndices)
 	}
@@ -282,9 +290,6 @@ func (f fakeSAT) RunPlatformStress(_ context.Context, _ string, _ platform.Platf
 func (f fakeSAT) RunNCCLTests(_ context.Context, _ string, _ func(string)) (string, error) {
 	return "", nil
 }
-func (f fakeSAT) RunHPL(_ context.Context, _ string, _ platform.HPLOptions, _ func(string)) (string, *platform.HPLResult, error) {
-	return "", nil, nil
-}
 
 func TestNetworkStatusFormatsInterfacesAndRoute(t *testing.T) {
 	t.Parallel()
@@ -545,8 +550,6 @@ func TestActionResultsUseFallbackBody(t *testing.T) {
 }
 
 func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -583,8 +586,6 @@ func TestExportSupportBundleResultMentionsUnmountedUSB(t *testing.T) {
 }
 
 func TestExportSupportBundleResultDoesNotPretendSuccessOnError(t *testing.T) {
-	t.Parallel()
-
 	tmp := t.TempDir()
 	oldExportDir := DefaultExportDir
 	DefaultExportDir = tmp
@@ -646,8 +647,6 @@ func TestRunNvidiaAcceptancePackResult(t *testing.T) {
 }
 
 func TestRunSATDefaultsToExportDir(t *testing.T) {
-	t.Parallel()
-
 	oldSATBaseDir := DefaultSATBaseDir
 	DefaultSATBaseDir = "/tmp/export/bee-sat"
 	t.Cleanup(func() { DefaultSATBaseDir = oldSATBaseDir })

audit/internal/app/support_bundle.go

@@ -22,6 +22,8 @@ var supportBundleServices = []string{
 	"bee-selfheal.service",
 	"bee-selfheal.timer",
 	"bee-sshsetup.service",
+	"nvidia-dcgm.service",
+	"nvidia-fabricmanager.service",
 }
 
 var supportBundleCommands = []struct {
@@ -48,13 +50,50 @@ else
 fi
 `}},
 	{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
+	{name: "system/nvidia-smi-topo.txt", cmd: []string{"sh", "-c", `
+if command -v nvidia-smi >/dev/null 2>&1; then
+  nvidia-smi topo -m 2>&1 || true
+else
+  echo "nvidia-smi not found"
+fi
+`}},
+	{name: "system/systemctl-nvidia-units.txt", cmd: []string{"sh", "-c", `
+if ! command -v systemctl >/dev/null 2>&1; then
+  echo "systemctl not found"
+  exit 0
+fi
+echo "=== unit files ==="
+systemctl list-unit-files --no-pager --all 'nvidia*' 'fabric*' 2>&1 || true
+echo
+echo "=== active units ==="
+systemctl list-units --no-pager --all 'nvidia*' 'fabric*' 2>&1 || true
+echo
+echo "=== failed units ==="
+systemctl --failed --no-pager 2>&1 | grep -iE 'nvidia|fabric' || echo "no failed nvidia/fabric units"
+`}},
+	{name: "system/fabric-manager-paths.txt", cmd: []string{"sh", "-c", `
+for candidate in \
+  /usr/bin/nvidia-fabricmanager \
+  /usr/bin/nv-fabricmanager \
+  /usr/bin/nvidia-fabricmanagerd \
+  /usr/bin/nvlsm; do
+  if [ -e "$candidate" ]; then
+    echo "=== $candidate ==="
+    ls -l "$candidate" 2>&1 || true
+    echo
+  fi
+done
+if ! ls /usr/bin/nvidia-fabricmanager /usr/bin/nv-fabricmanager /usr/bin/nvidia-fabricmanagerd /usr/bin/nvlsm >/dev/null 2>&1; then
+  echo "no fabric manager binaries found"
+fi
+`}},
 	{name: "system/lspci-nvidia-bridges-vv.txt", cmd: []string{"sh", "-c", `
 if ! command -v lspci >/dev/null 2>&1; then
   echo "lspci not found"
   exit 0
 fi
 found=0
-for gpu in $(lspci -Dn | awk '$3 ~ /^10de:/ {print $1}'); do
+	for gpu in $(lspci -Dn | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ {print $1}'); do
   found=1
   echo "=== GPU $gpu ==="
   lspci -s "$gpu" -vv 2>&1 || true
@@ -73,8 +112,13 @@ fi
 	{name: "system/pcie-nvidia-link.txt", cmd: []string{"sh", "-c", `
 for d in /sys/bus/pci/devices/*/; do
   vendor=$(cat "$d/vendor" 2>/dev/null)
-  [ "$vendor" = "0x10de" ] || continue
-  dev=$(basename "$d")
+	  [ "$vendor" = "0x10de" ] || continue
+	  class=$(cat "$d/class" 2>/dev/null)
+	  case "$class" in
+	    0x030000|0x030200) ;;
+	    *) continue ;;
+	  esac
+	  dev=$(basename "$d")
   echo "=== $dev ==="
   for f in current_link_speed current_link_width max_link_speed max_link_width; do
     printf "  %-22s %s\n" "$f" "$(cat "$d/$f" 2>/dev/null)"
@@ -190,9 +234,13 @@ var supportBundleOptionalFiles = []struct {
 }{
 	{name: "system/kern.log", src: "/var/log/kern.log"},
 	{name: "system/syslog.txt", src: "/var/log/syslog"},
+	{name: "system/fabricmanager.log", src: "/var/log/fabricmanager.log"},
+	{name: "system/nvlsm.log", src: "/var/log/nvlsm.log"},
+	{name: "system/fabricmanager/fabricmanager.log", src: "/var/log/fabricmanager/fabricmanager.log"},
+	{name: "system/fabricmanager/nvlsm.log", src: "/var/log/fabricmanager/nvlsm.log"},
 }
 
-const supportBundleGlob = "bee-support-*.tar.gz"
+const supportBundleGlob = "????-??-?? (BEE-SP*)*.tar.gz"
 
 func BuildSupportBundle(exportDir string) (string, error) {
 	exportDir = strings.TrimSpace(exportDir)
@@ -206,9 +254,14 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	host := sanitizeFilename(hostnameOr("unknown"))
-	ts := time.Now().UTC().Format("20060102-150405")
-	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s", host, ts))
+	now := time.Now().UTC()
+	date := now.Format("2006-01-02")
+	tod := now.Format("150405")
+	ver := bundleVersion()
+	model := serverModelForBundle()
+	sn := serverSerialForBundle()
+
+	stageRoot := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-stage-%s-%s", sanitizeFilename(hostnameOr("unknown")), now.Format("20060102-150405")))
 	if err := os.MkdirAll(stageRoot, 0755); err != nil {
 		return "", err
 	}
@@ -240,7 +293,8 @@ func BuildSupportBundle(exportDir string) (string, error) {
 		return "", err
 	}
 
-	archivePath := filepath.Join(os.TempDir(), fmt.Sprintf("bee-support-%s-%s.tar.gz", host, ts))
+	archiveName := fmt.Sprintf("%s (BEE-SP v%s) %s %s %s.tar.gz", date, ver, model, sn, tod)
+	archivePath := filepath.Join(os.TempDir(), archiveName)
 	if err := createSupportTarGz(archivePath, stageRoot); err != nil {
 		return "", err
 	}
@@ -397,6 +451,60 @@ func writeManifest(dst, exportDir, stageRoot string) error {
 	return os.WriteFile(dst, []byte(body.String()), 0644)
 }
 
+func bundleVersion() string {
+	v := buildVersion()
+	v = strings.TrimPrefix(v, "v")
+	v = strings.TrimPrefix(v, "V")
+	if v == "" || v == "unknown" {
+		return "0.0"
+	}
+	return v
+}
+
+func serverModelForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Product Name" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return strings.ReplaceAll(val, " ", "_")
+		}
+	}
+	return "unknown"
+}
+
+func serverSerialForBundle() string {
+	raw, err := exec.Command("dmidecode", "-t", "1").Output()
+	if err != nil {
+		return "unknown"
+	}
+	for _, line := range strings.Split(string(raw), "\n") {
+		line = strings.TrimSpace(line)
+		key, val, ok := strings.Cut(line, ": ")
+		if !ok {
+			continue
+		}
+		if strings.TrimSpace(key) == "Serial Number" {
+			val = strings.TrimSpace(val)
+			if val == "" {
+				return "unknown"
+			}
+			return val
+		}
+	}
+	return "unknown"
+}
+
 func buildVersion() string {
 	raw, err := exec.Command("bee", "version").CombinedOutput()
 	if err != nil {

audit/internal/collector/nic_mellanox.go

@@ -179,11 +179,3 @@ func commandOutputWithTimeout(timeout time.Duration, name string, args ...string
 	defer cancel()
 	return exec.CommandContext(ctx, name, args...).Output()
 }
-
-func interfaceHasCarrier(iface string) bool {
-	raw, err := readNetCarrierFile(iface)
-	if err != nil {
-		return false
-	}
-	return strings.TrimSpace(raw) == "1"
-}

audit/internal/collector/nic_telemetry.go

@@ -58,12 +58,10 @@ func enrichPCIeWithNICTelemetry(devs []schema.HardwarePCIeDevice) []schema.Hardw
 			}
 		}
 
-		if interfaceHasCarrier(iface) {
-			if out, err := ethtoolModuleQuery(iface); err == nil {
-				if injectSFPDOMTelemetry(&devs[i], out) {
-					enriched++
-					continue
-				}
+		if out, err := ethtoolModuleQuery(iface); err == nil {
+			if injectSFPDOMTelemetry(&devs[i], out) {
+				enriched++
+				continue
 			}
 		}
 		if len(devs[i].MacAddresses) > 0 || devs[i].Firmware != nil {
@@ -115,8 +113,38 @@ func injectSFPDOMTelemetry(dev *schema.HardwarePCIeDevice, raw string) bool {
 		}
 		key := strings.ToLower(strings.TrimSpace(trimmed[:idx]))
 		val := strings.TrimSpace(trimmed[idx+1:])
+		if val == "" || strings.EqualFold(val, "not supported") || strings.EqualFold(val, "unknown") {
+			continue
+		}
 
 		switch {
+		case key == "identifier":
+			s := parseSFPIdentifier(val)
+			dev.SFPIdentifier = &s
+			t := true
+			dev.SFPPresent = &t
+			changed = true
+		case key == "connector":
+			s := parseSFPConnector(val)
+			dev.SFPConnector = &s
+			changed = true
+		case key == "vendor name":
+			s := strings.TrimSpace(val)
+			dev.SFPVendor = &s
+			changed = true
+		case key == "vendor pn":
+			s := strings.TrimSpace(val)
+			dev.SFPPartNumber = &s
+			changed = true
+		case key == "vendor sn":
+			s := strings.TrimSpace(val)
+			dev.SFPSerialNumber = &s
+			changed = true
+		case strings.Contains(key, "laser wavelength"):
+			if f, ok := firstFloat(val); ok {
+				dev.SFPWavelengthNM = &f
+				changed = true
+			}
 		case strings.Contains(key, "module temperature"):
 			if f, ok := firstFloat(val); ok {
 				dev.SFPTemperatureC = &f
@@ -147,12 +175,61 @@ func injectSFPDOMTelemetry(dev *schema.HardwarePCIeDevice, raw string) bool {
 	return changed
 }
 
+// parseSFPIdentifier extracts the human-readable transceiver type from the
+// raw ethtool identifier line, e.g. "0x03 (SFP)" → "SFP".
+func parseSFPIdentifier(val string) string {
+	if s := extractParens(val); s != "" {
+		return s
+	}
+	return val
+}
+
+// parseSFPConnector extracts the connector type from the raw ethtool line,
+// e.g. "0x07 (LC)" → "LC".
+func parseSFPConnector(val string) string {
+	if s := extractParens(val); s != "" {
+		return s
+	}
+	return val
+}
+
+var parenRe = regexp.MustCompile(`\(([^)]+)\)`)
+
+func extractParens(s string) string {
+	m := parenRe.FindStringSubmatch(s)
+	if len(m) < 2 {
+		return ""
+	}
+	return strings.TrimSpace(m[1])
+}
+
 func parseSFPDOM(raw string) map[string]any {
 	dev := schema.HardwarePCIeDevice{}
 	if !injectSFPDOMTelemetry(&dev, raw) {
 		return map[string]any{}
 	}
 	out := map[string]any{}
+	if dev.SFPPresent != nil {
+		out["sfp_present"] = *dev.SFPPresent
+	}
+	if dev.SFPIdentifier != nil {
+		out["sfp_identifier"] = *dev.SFPIdentifier
+	}
+	if dev.SFPConnector != nil {
+		out["sfp_connector"] = *dev.SFPConnector
+	}
+	if dev.SFPVendor != nil {
+		out["sfp_vendor"] = *dev.SFPVendor
+	}
+	if dev.SFPPartNumber != nil {
+		out["sfp_part_number"] = *dev.SFPPartNumber
+	}
+	if dev.SFPSerialNumber != nil {
+		out["sfp_serial_number"] = *dev.SFPSerialNumber
+	}
+	if dev.SFPWavelengthNM != nil {
+		out["sfp_wavelength_nm"] = *dev.SFPWavelengthNM
+	}
 	if dev.SFPTemperatureC != nil {
 		out["sfp_temperature_c"] = *dev.SFPTemperatureC
 	}

audit/internal/collector/nic_telemetry_test.go

@@ -122,10 +122,7 @@ func TestEnrichPCIeWithNICTelemetrySkipsModuleQueryWithoutCarrier(t *testing.T)
 	readNetAddressFile = func(string) (string, error) { return "aa:bb:cc:dd:ee:ff", nil }
 	readNetCarrierFile = func(string) (string, error) { return "0", nil }
 	ethtoolInfoQuery = func(string) (string, error) { return "", fmt.Errorf("skip firmware") }
-	ethtoolModuleQuery = func(string) (string, error) {
-		t.Fatal("ethtool -m should not be called without carrier")
-		return "", nil
-	}
+	ethtoolModuleQuery = func(string) (string, error) { return "", fmt.Errorf("no module") }
 
 	class := "EthernetController"
 	bdf := "0000:18:00.0"

audit/internal/collector/nvidia.go

@@ -15,6 +15,7 @@ const nvidiaVendorID = 0x10de
 type nvidiaGPUInfo struct {
 	Index              int
 	BDF                string
+	Name               string
 	Serial             string
 	VBIOS              string
 	TemperatureC       *float64
@@ -73,6 +74,9 @@ func enrichPCIeWithNVIDIAData(devs []schema.HardwarePCIeDevice, gpuByBDF map[str
 			continue
 		}
 
+		if v := strings.TrimSpace(info.Name); v != "" {
+			devs[i].Model = &v
+		}
 		if v := strings.TrimSpace(info.Serial); v != "" {
 			devs[i].SerialNumber = &v
 		}
@@ -99,7 +103,7 @@ func enrichPCIeWithNVIDIAData(devs []schema.HardwarePCIeDevice, gpuByBDF map[str
 func queryNVIDIAGPUs() (map[string]nvidiaGPUInfo, error) {
 	out, err := exec.Command(
 		"nvidia-smi",
-		"--query-gpu=index,pci.bus_id,serial,vbios_version,temperature.gpu,power.draw,ecc.errors.uncorrected.aggregate.total,ecc.errors.corrected.aggregate.total,clocks_throttle_reasons.hw_slowdown,pcie.link.gen.current,pcie.link.gen.max,pcie.link.width.current,pcie.link.width.max",
+		"--query-gpu=index,pci.bus_id,name,serial,vbios_version,temperature.gpu,power.draw,ecc.errors.uncorrected.aggregate.total,ecc.errors.corrected.aggregate.total,clocks_throttle_reasons.hw_slowdown,pcie.link.gen.current,pcie.link.gen.max,pcie.link.width.current,pcie.link.width.max",
 		"--format=csv,noheader,nounits",
 	).Output()
 	if err != nil {
@@ -123,8 +127,8 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		if len(rec) == 0 {
 			continue
 		}
-		if len(rec) < 13 {
-			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 13", len(rec))
+		if len(rec) < 14 {
+			return nil, fmt.Errorf("unexpected nvidia-smi columns: got %d, want 14", len(rec))
 		}
 
 		bdf := normalizePCIeBDF(rec[1])
@@ -135,17 +139,18 @@ func parseNVIDIASMIQuery(raw string) (map[string]nvidiaGPUInfo, error) {
 		info := nvidiaGPUInfo{
 			Index:              parseRequiredInt(rec[0]),
 			BDF:                bdf,
-			Serial:             strings.TrimSpace(rec[2]),
-			VBIOS:              strings.TrimSpace(rec[3]),
-			TemperatureC:       parseMaybeFloat(rec[4]),
-			PowerW:             parseMaybeFloat(rec[5]),
-			ECCUncorrected:     parseMaybeInt64(rec[6]),
-			ECCCorrected:       parseMaybeInt64(rec[7]),
-			HWSlowdown:         parseMaybeBool(rec[8]),
-			PCIeLinkGenCurrent: parseMaybeInt(rec[9]),
-			PCIeLinkGenMax:     parseMaybeInt(rec[10]),
-			PCIeLinkWidthCur:   parseMaybeInt(rec[11]),
-			PCIeLinkWidthMax:   parseMaybeInt(rec[12]),
+			Name:               strings.TrimSpace(rec[2]),
+			Serial:             strings.TrimSpace(rec[3]),
+			VBIOS:              strings.TrimSpace(rec[4]),
+			TemperatureC:       parseMaybeFloat(rec[5]),
+			PowerW:             parseMaybeFloat(rec[6]),
+			ECCUncorrected:     parseMaybeInt64(rec[7]),
+			ECCCorrected:       parseMaybeInt64(rec[8]),
+			HWSlowdown:         parseMaybeBool(rec[9]),
+			PCIeLinkGenCurrent: parseMaybeInt(rec[10]),
+			PCIeLinkGenMax:     parseMaybeInt(rec[11]),
+			PCIeLinkWidthCur:   parseMaybeInt(rec[12]),
+			PCIeLinkWidthMax:   parseMaybeInt(rec[13]),
 		}
 		result[bdf] = info
 	}

audit/internal/collector/nvidia_test.go

@@ -6,7 +6,7 @@ import (
 )
 
 func TestParseNVIDIASMIQuery(t *testing.T) {
-	raw := "0, 00000000:65:00.0, GPU-SERIAL-1, 96.00.1F.00.02, 54, 210.33, 0, 5, Not Active, 4, 4, 16, 16\n"
+	raw := "0, 00000000:65:00.0, NVIDIA H100 80GB HBM3, GPU-SERIAL-1, 96.00.1F.00.02, 54, 210.33, 0, 5, Not Active, 4, 4, 16, 16\n"
 	byBDF, err := parseNVIDIASMIQuery(raw)
 	if err != nil {
 		t.Fatalf("parse failed: %v", err)
@@ -16,6 +16,9 @@ func TestParseNVIDIASMIQuery(t *testing.T) {
 	if !ok {
 		t.Fatalf("gpu by normalized bdf not found")
 	}
+	if gpu.Name != "NVIDIA H100 80GB HBM3" {
+		t.Fatalf("name: got %q", gpu.Name)
+	}
 	if gpu.Serial != "GPU-SERIAL-1" {
 		t.Fatalf("serial: got %q", gpu.Serial)
 	}

audit/internal/collector/pcie.go

@@ -2,6 +2,7 @@ package collector
 
 import (
 	"bee/audit/internal/schema"
+	"fmt"
 	"log/slog"
 	"os/exec"
 	"strconv"
@@ -79,6 +80,25 @@ func shouldIncludePCIeDevice(class, vendor, device string) bool {
 		}
 	}
 
+	// Exclude BMC/management virtual VGA adapters — these are firmware video chips,
+	// not real GPUs, and pollute the GPU inventory (e.g. iBMC, iDRAC, iLO VGA).
+	if strings.Contains(c, "vga") || strings.Contains(c, "display") || strings.Contains(c, "3d") {
+		bmcPatterns := []string{
+			"management system chip",
+			"management controller",
+			"ibmc",
+			"idrac",
+			"ilo vga",
+			"aspeed",
+			"matrox",
+		}
+		for _, bad := range bmcPatterns {
+			if strings.Contains(d, bad) {
+				return false
+			}
+		}
+	}
+
 	if strings.Contains(v, "advanced micro devices") || strings.Contains(v, "[amd]") {
 		internalAMDPatterns := []string{
 			"dummy function",
@@ -153,6 +173,9 @@ func parseLspciDevice(fields map[string]string) schema.HardwarePCIeDevice {
 
 	// SVendor/SDevice available but not in schema — skip
 
+	// Warn if PCIe link is running below its maximum negotiated speed.
+	applyPCIeLinkSpeedWarning(&dev)
+
 	return dev
 }
 
@@ -222,6 +245,41 @@ func readPCIStringAttribute(bdf, attribute string) (string, bool) {
 	return value, true
 }
 
+// applyPCIeLinkSpeedWarning sets the device status to Warning if the current PCIe link
+// speed is below the maximum negotiated speed supported by both ends.
+func applyPCIeLinkSpeedWarning(dev *schema.HardwarePCIeDevice) {
+	if dev.LinkSpeed == nil || dev.MaxLinkSpeed == nil {
+		return
+	}
+	if pcieLinkSpeedRank(*dev.LinkSpeed) < pcieLinkSpeedRank(*dev.MaxLinkSpeed) {
+		warn := statusWarning
+		dev.Status = &warn
+		desc := fmt.Sprintf("PCIe link speed degraded: running at %s, capable of %s", *dev.LinkSpeed, *dev.MaxLinkSpeed)
+		dev.ErrorDescription = &desc
+	}
+}
+
+// pcieLinkSpeedRank returns a numeric rank for a normalized Gen string (e.g. "Gen4" → 4).
+// Returns 0 for unrecognised values so comparisons fail safe.
+func pcieLinkSpeedRank(gen string) int {
+	switch gen {
+	case "Gen1":
+		return 1
+	case "Gen2":
+		return 2
+	case "Gen3":
+		return 3
+	case "Gen4":
+		return 4
+	case "Gen5":
+		return 5
+	case "Gen6":
+		return 6
+	default:
+		return 0
+	}
+}
+
 func normalizePCILinkSpeed(raw string) string {
 	raw = strings.TrimSpace(strings.ToLower(raw))
 	switch {

audit/internal/collector/pcie_filter_test.go

@@ -1,6 +1,7 @@
 package collector
 
 import (
+	"bee/audit/internal/schema"
 	"encoding/json"
 	"strings"
 	"testing"
@@ -29,6 +30,8 @@ func TestShouldIncludePCIeDevice(t *testing.T) {
 		{name: "raid", class: "RAID bus controller", want: true},
 		{name: "nvme", class: "Non-Volatile memory controller", want: true},
 		{name: "vga", class: "VGA compatible controller", want: true},
+		{name: "ibmc vga", class: "VGA compatible controller", vendor: "Huawei Technologies Co., Ltd.", device: "Hi171x Series [iBMC Intelligent Management system chip w/VGA support]", want: false},
+		{name: "aspeed vga", class: "VGA compatible controller", vendor: "ASPEED Technology, Inc.", device: "ASPEED Graphics Family", want: false},
 		{name: "other encryption controller", class: "Encryption controller", vendor: "Intel Corporation", device: "QuickAssist", want: true},
 	}
 
@@ -139,3 +142,77 @@ func TestNormalizePCILinkSpeed(t *testing.T) {
 		}
 	}
 }
+
+func TestApplyPCIeLinkSpeedWarning(t *testing.T) {
+	ptr := func(s string) *string { return &s }
+
+	tests := []struct {
+		name        string
+		linkSpeed   *string
+		maxSpeed    *string
+		wantWarning bool
+		wantGenIn   string // substring expected in ErrorDescription when warning
+	}{
+		{
+			name:        "degraded Gen1 vs Gen5",
+			linkSpeed:   ptr("Gen1"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: true,
+			wantGenIn:   "Gen1",
+		},
+		{
+			name:        "at max Gen5",
+			linkSpeed:   ptr("Gen5"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: false,
+		},
+		{
+			name:        "degraded Gen4 vs Gen5",
+			linkSpeed:   ptr("Gen4"),
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: true,
+			wantGenIn:   "Gen4",
+		},
+		{
+			name:        "missing current speed — no warning",
+			linkSpeed:   nil,
+			maxSpeed:    ptr("Gen5"),
+			wantWarning: false,
+		},
+		{
+			name:        "missing max speed — no warning",
+			linkSpeed:   ptr("Gen1"),
+			maxSpeed:    nil,
+			wantWarning: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dev := schema.HardwarePCIeDevice{}
+			ok := statusOK
+			dev.Status = &ok
+			dev.LinkSpeed = tt.linkSpeed
+			dev.MaxLinkSpeed = tt.maxSpeed
+
+			applyPCIeLinkSpeedWarning(&dev)
+
+			gotWarn := dev.Status != nil && *dev.Status == statusWarning
+			if gotWarn != tt.wantWarning {
+				t.Fatalf("wantWarning=%v gotWarning=%v (status=%v)", tt.wantWarning, gotWarn, dev.Status)
+			}
+			if tt.wantWarning {
+				if dev.ErrorDescription == nil {
+					t.Fatal("expected ErrorDescription to be set")
+				}
+				if !strings.Contains(*dev.ErrorDescription, tt.wantGenIn) {
+					t.Fatalf("ErrorDescription %q does not contain %q", *dev.ErrorDescription, tt.wantGenIn)
+				}
+			} else {
+				if dev.ErrorDescription != nil {
+					t.Fatalf("unexpected ErrorDescription: %s", *dev.ErrorDescription)
+				}
+			}
+		})
+	}
+}

audit/internal/platform/benchmark_report.go

@@ -2,25 +2,15 @@ package platform
 
 import (
 	"fmt"
-	"os"
-	"path/filepath"
-	"regexp"
 	"strings"
 	"time"
 )
 
 func renderBenchmarkReport(result NvidiaBenchmarkResult) string {
-	return renderBenchmarkReportWithCharts(result, nil)
+	return renderBenchmarkReportWithCharts(result)
 }
 
-type benchmarkReportChart struct {
-	Title   string
-	Content string
-}
-
-var ansiEscapePattern = regexp.MustCompile(`\x1b\[[0-9;]*m`)
-
-func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benchmarkReportChart) string {
+func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 	var b strings.Builder
 
 	// ── Header ────────────────────────────────────────────────────────────────
@@ -58,11 +48,22 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		fmt.Fprintf(&b, "**GPU(s):** %s  \n", strings.Join(parts, ", "))
 	}
 	fmt.Fprintf(&b, "**Profile:** %s  \n", result.BenchmarkProfile)
-	fmt.Fprintf(&b, "**App version:** %s  \n", result.BenchmarkVersion)
+	fmt.Fprintf(&b, "**Benchmark version:** %s  \n", result.BenchmarkVersion)
 	fmt.Fprintf(&b, "**Generated:** %s  \n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
-	if result.ParallelGPUs {
+	if result.RampStep > 0 && result.RampTotal > 0 {
+		fmt.Fprintf(&b, "**Ramp-up step:** %d of %d  \n", result.RampStep, result.RampTotal)
+		if result.RampRunID != "" {
+			fmt.Fprintf(&b, "**Ramp-up run ID:** %s  \n", result.RampRunID)
+		}
+	} else if result.ParallelGPUs {
 		fmt.Fprintf(&b, "**Mode:** parallel (all GPUs simultaneously)  \n")
 	}
+	if result.ScalabilityScore > 0 {
+		fmt.Fprintf(&b, "**Scalability score:** %.1f%%  \n", result.ScalabilityScore)
+	}
+	if result.PlatformPowerScore > 0 {
+		fmt.Fprintf(&b, "**Platform power score:** %.1f%%  \n", result.PlatformPowerScore)
+	}
 	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
 	b.WriteString("\n")
 
@@ -83,34 +84,139 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		b.WriteString("\n")
 	}
 
-	// ── Scorecard table ───────────────────────────────────────────────────────
-	b.WriteString("## Scorecard\n\n")
-	b.WriteString("| GPU | Status | Composite | Compute | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
-	b.WriteString("|-----|--------|-----------|---------|-------------|---------------|-----------------|-----------|-------------|\n")
+	// ── Balanced Scorecard ────────────────────────────────────────────────────
+	b.WriteString("## Balanced Scorecard\n\n")
+
+	// Perspective 1: Compatibility — hard stops
+	b.WriteString("### 1. Compatibility\n\n")
+	b.WriteString("| GPU | Thermal throttle | Fan duty at throttle | ECC uncorr | Status |\n")
+	b.WriteString("|-----|------------------|----------------------|------------|--------|\n")
 	for _, gpu := range result.GPUs {
-		name := strings.TrimSpace(gpu.Name)
-		if name == "" {
-			name = "Unknown"
+		thermalThrottle := "-"
+		if gpu.Scores.ThermalThrottlePct > 0 {
+			thermalThrottle = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
 		}
-		interconnect := "-"
-		if gpu.Scores.InterconnectScore > 0 {
-			interconnect = fmt.Sprintf("%.1f", gpu.Scores.InterconnectScore)
+		fanAtThrottle := "-"
+		if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && gpu.Scores.ThermalThrottlePct > 0 {
+			fanAtThrottle = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
+		}
+		ecc := "-"
+		if gpu.ECC.Uncorrected > 0 {
+			ecc = fmt.Sprintf("⛔ %d", gpu.ECC.Uncorrected)
+		}
+		compatStatus := "✓ OK"
+		if gpu.ECC.Uncorrected > 0 || (gpu.Scores.ThermalThrottlePct > 0 && result.Cooling != nil && result.Cooling.FanDutyCycleAvailable && result.Cooling.P95FanDutyCyclePct < 95) {
+			compatStatus = "⛔ HARD STOP"
+		}
+		fmt.Fprintf(&b, "| GPU %d | %s | %s | %s | %s |\n",
+			gpu.Index, thermalThrottle, fanAtThrottle, ecc, compatStatus)
+	}
+	b.WriteString("\n")
+
+	// Perspective 2: Thermal headroom
+	b.WriteString("### 2. Thermal Headroom\n\n")
+	b.WriteString("| GPU | p95 temp | Slowdown limit | Shutdown limit | Headroom | Thermal throttle | Status |\n")
+	b.WriteString("|-----|----------|----------------|----------------|----------|------------------|--------|\n")
+	for _, gpu := range result.GPUs {
+		shutdownTemp := gpu.ShutdownTempC
+		if shutdownTemp <= 0 {
+			shutdownTemp = 90
+		}
+		slowdownTemp := gpu.SlowdownTempC
+		if slowdownTemp <= 0 {
+			slowdownTemp = 80
+		}
+		headroom := gpu.Scores.TempHeadroomC
+		thermalStatus := "✓ OK"
+		switch {
+		case headroom < 10:
+			thermalStatus = "⛔ CRITICAL"
+		case gpu.Steady.P95TempC >= slowdownTemp:
+			thermalStatus = "⚠ WARNING"
+		}
+		throttlePct := "-"
+		if gpu.Scores.ThermalThrottlePct > 0 {
+			throttlePct = fmt.Sprintf("%.1f%%", gpu.Scores.ThermalThrottlePct)
+		}
+		fmt.Fprintf(&b, "| GPU %d | %.1f°C | %.0f°C | %.0f°C | %.1f°C | %s | %s |\n",
+			gpu.Index, gpu.Steady.P95TempC, slowdownTemp, shutdownTemp, headroom, throttlePct, thermalStatus)
+	}
+	b.WriteString("\n")
+
+	// Perspective 3: Power delivery
+	b.WriteString("### 3. Power Delivery\n\n")
+	b.WriteString("| GPU | Power cap throttle | Power stability | Fan duty (p95) | Status |\n")
+	b.WriteString("|-----|-------------------|-----------------|----------------|--------|\n")
+	for _, gpu := range result.GPUs {
+		powerCap := "-"
+		if gpu.Scores.PowerCapThrottlePct > 0 {
+			powerCap = fmt.Sprintf("%.1f%%", gpu.Scores.PowerCapThrottlePct)
+		}
+		fanDuty := "-"
+		if result.Cooling != nil && result.Cooling.FanDutyCycleAvailable {
+			fanDuty = fmt.Sprintf("%.0f%%", result.Cooling.P95FanDutyCyclePct)
+		}
+		powerStatus := "✓ OK"
+		if gpu.Scores.PowerCapThrottlePct > 5 {
+			powerStatus = "⚠ POWER LIMITED"
+		}
+		fmt.Fprintf(&b, "| GPU %d | %s | %.1f | %s | %s |\n",
+			gpu.Index, powerCap, gpu.Scores.PowerSustainScore, fanDuty, powerStatus)
+	}
+	b.WriteString("\n")
+
+	// Perspective 4: Performance
+	b.WriteString("### 4. Performance\n\n")
+	b.WriteString("| GPU | Compute TOPS | Synthetic | Mixed | Mixed Eff. | TOPS/SM/GHz |\n")
+	b.WriteString("|-----|--------------|-----------|-------|------------|-------------|\n")
+	for _, gpu := range result.GPUs {
+		synthetic := "-"
+		if gpu.Scores.SyntheticScore > 0 {
+			synthetic = fmt.Sprintf("%.2f", gpu.Scores.SyntheticScore)
+		}
+		mixed := "-"
+		if gpu.Scores.MixedScore > 0 {
+			mixed = fmt.Sprintf("%.2f", gpu.Scores.MixedScore)
+		}
+		mixedEff := "-"
+		if gpu.Scores.MixedEfficiency > 0 {
+			mixedEff = fmt.Sprintf("%.1f%%", gpu.Scores.MixedEfficiency*100)
 		}
 		topsPerSM := "-"
 		if gpu.Scores.TOPSPerSMPerGHz > 0 {
 			topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
 		}
-		fmt.Fprintf(&b, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %.1f | %.1f | %.1f | %s |\n",
-			gpu.Index, name,
-			gpu.Status,
-			gpu.Scores.CompositeScore,
-			gpu.Scores.ComputeScore,
-			topsPerSM,
-			gpu.Scores.PowerSustainScore,
-			gpu.Scores.ThermalSustainScore,
-			gpu.Scores.StabilityScore,
-			interconnect,
-		)
+		fmt.Fprintf(&b, "| GPU %d | **%.2f** | %s | %s | %s | %s |\n",
+			gpu.Index, gpu.Scores.CompositeScore, synthetic, mixed, mixedEff, topsPerSM)
+	}
+	if len(result.PerformanceRampSteps) > 0 {
+		fmt.Fprintf(&b, "\n**Platform power score (scalability):** %.1f%%\n", result.PlatformPowerScore)
+	}
+	b.WriteString("\n")
+
+	// Perspective 5: Anomaly flags
+	b.WriteString("### 5. Anomalies\n\n")
+	b.WriteString("| GPU | ECC corrected | Sync boost throttle | Power instability | Thermal instability |\n")
+	b.WriteString("|-----|---------------|---------------------|-------------------|---------------------|\n")
+	for _, gpu := range result.GPUs {
+		eccCorr := "-"
+		if gpu.ECC.Corrected > 0 {
+			eccCorr = fmt.Sprintf("⚠ %d", gpu.ECC.Corrected)
+		}
+		syncBoost := "-"
+		if gpu.Scores.SyncBoostThrottlePct > 0 {
+			syncBoost = fmt.Sprintf("%.1f%%", gpu.Scores.SyncBoostThrottlePct)
+		}
+		powerVar := "OK"
+		if gpu.Scores.PowerSustainScore < 70 {
+			powerVar = "⚠ unstable"
+		}
+		thermalVar := "OK"
+		if gpu.Scores.ThermalSustainScore < 70 {
+			thermalVar = "⚠ unstable"
+		}
+		fmt.Fprintf(&b, "| GPU %d | %s | %s | %s | %s |\n",
+			gpu.Index, eccCorr, syncBoost, powerVar, thermalVar)
 	}
 	b.WriteString("\n")
 
@@ -139,20 +245,66 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		if gpu.PowerLimitW > 0 {
 			fmt.Fprintf(&b, "- **Power limit:** %.0f W (default %.0f W)\n", gpu.PowerLimitW, gpu.DefaultPowerLimitW)
 		}
+		if gpu.PowerLimitDerated {
+			fmt.Fprintf(&b, "- **Power limit derating:** active (reduced limit %.0f W)\n", gpu.PowerLimitW)
+		}
+		if gpu.CalibratedPeakPowerW > 0 {
+			if gpu.CalibratedPeakTempC > 0 {
+				fmt.Fprintf(&b, "- **Calibrated peak power:** %.0f W p95 at %.1f °C p95\n", gpu.CalibratedPeakPowerW, gpu.CalibratedPeakTempC)
+			} else {
+				fmt.Fprintf(&b, "- **Calibrated peak power:** %.0f W p95\n", gpu.CalibratedPeakPowerW)
+			}
+		}
 		if gpu.LockedGraphicsClockMHz > 0 {
 			fmt.Fprintf(&b, "- **Locked clocks:** GPU %.0f MHz / Mem %.0f MHz\n", gpu.LockedGraphicsClockMHz, gpu.LockedMemoryClockMHz)
 		}
 		b.WriteString("\n")
 
 		// Steady-state telemetry
-		fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
-		b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
-		fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
-		fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
-		fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
-		fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
-		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
-		b.WriteString("\n")
+		if benchmarkTelemetryAvailable(gpu.Steady) {
+			fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
+			b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
+			fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
+			fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
+			fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
+			fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
+			fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
+			b.WriteString("\n")
+		} else {
+			b.WriteString("**Steady-state telemetry:** unavailable\n\n")
+		}
+
+		// Per-precision stability phases.
+		if len(gpu.PrecisionSteady) > 0 {
+			b.WriteString("**Per-precision stability:**\n\n")
+			b.WriteString("| Precision | Status | Clock CV | Power CV | Clock Drift | ECC corr | ECC uncorr |\n|-----------|--------|----------|----------|-------------|----------|------------|\n")
+			for _, p := range gpu.PrecisionSteady {
+				eccCorr := "—"
+				eccUncorr := "—"
+				if !p.ECC.IsZero() {
+					eccCorr = fmt.Sprintf("%d", p.ECC.Corrected)
+					eccUncorr = fmt.Sprintf("%d", p.ECC.Uncorrected)
+				}
+				status := p.Status
+				if strings.TrimSpace(status) == "" {
+					status = "OK"
+				}
+				fmt.Fprintf(&b, "| %s | %s | %.1f%% | %.1f%% | %.1f%% | %s | %s |\n",
+					p.Precision, status, p.Steady.ClockCVPct, p.Steady.PowerCVPct, p.Steady.ClockDriftPct,
+					eccCorr, eccUncorr)
+			}
+			b.WriteString("\n")
+		} else {
+			// Legacy: show combined-window variance.
+			fmt.Fprintf(&b, "**Clock/power variance (combined window):** clock CV %.1f%% · power CV %.1f%% · clock drift %.1f%%\n\n",
+				gpu.Steady.ClockCVPct, gpu.Steady.PowerCVPct, gpu.Steady.ClockDriftPct)
+		}
+
+		// ECC summary
+		if !gpu.ECC.IsZero() {
+			fmt.Fprintf(&b, "**ECC errors (total):** corrected=%d uncorrected=%d\n\n",
+				gpu.ECC.Corrected, gpu.ECC.Uncorrected)
+		}
 
 		// Throttle
 		throttle := formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec)
@@ -163,12 +315,14 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		// Precision results
 		if len(gpu.PrecisionResults) > 0 {
 			b.WriteString("**Precision results:**\n\n")
-			b.WriteString("| Precision | TOPS | Lanes | Iterations |\n|-----------|------|-------|------------|\n")
+			b.WriteString("| Precision | TOPS (raw) | Weight | TOPS (fp32-eq) | Lanes | Iterations |\n|-----------|------------|--------|----------------|-------|------------|\n")
 			for _, p := range gpu.PrecisionResults {
 				if p.Supported {
-					fmt.Fprintf(&b, "| %s | %.2f | %d | %d |\n", p.Name, p.TeraOpsPerSec, p.Lanes, p.Iterations)
+					weightStr := fmt.Sprintf("×%.3g", p.Weight)
+					fmt.Fprintf(&b, "| %s | %.2f | %s | %.2f | %d | %d |\n",
+						p.Name, p.TeraOpsPerSec, weightStr, p.WeightedTeraOpsPerSec, p.Lanes, p.Iterations)
 				} else {
-					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — |\n", p.Name)
+					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — | — | — |\n", p.Name)
 				}
 			}
 			b.WriteString("\n")
@@ -229,61 +383,54 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		}
 	}
 
-	// ── Terminal charts (steady-state only) ───────────────────────────────────
-	if len(charts) > 0 {
-		b.WriteString("## Steady-State Charts\n\n")
-		for _, chart := range charts {
-			content := strings.TrimSpace(stripANSIEscapeSequences(chart.Content))
-			if content == "" {
-				continue
+	// ── Cooling ───────────────────────────────────────────────────────────────
+	if cooling := result.Cooling; cooling != nil {
+		b.WriteString("## Cooling\n\n")
+		if cooling.Available {
+			b.WriteString("| Metric | Value |\n|--------|-------|\n")
+			fmt.Fprintf(&b, "| Average fan speed | %.0f RPM |\n", cooling.AvgFanRPM)
+			if cooling.FanDutyCycleAvailable {
+				fmt.Fprintf(&b, "| Average fan duty cycle | %.1f%% |\n", cooling.AvgFanDutyCyclePct)
+				fmt.Fprintf(&b, "| P95 fan duty cycle | %.1f%% |\n", cooling.P95FanDutyCyclePct)
+			} else {
+				b.WriteString("| Average fan duty cycle | N/A |\n")
+				b.WriteString("| P95 fan duty cycle | N/A |\n")
 			}
-			fmt.Fprintf(&b, "### %s\n\n```\n%s\n```\n\n", chart.Title, content)
+			b.WriteString("\n")
+		} else {
+			b.WriteString("Cooling telemetry unavailable.\n\n")
+		}
+		for _, note := range cooling.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		if len(cooling.Notes) > 0 {
+			b.WriteString("\n")
 		}
 	}
 
-	// ── Methodology ───────────────────────────────────────────────────────────
-	b.WriteString("## Methodology\n\n")
-	fmt.Fprintf(&b, "- Profile `%s` uses standardized baseline → warmup → steady-state → interconnect → cooldown phases.\n", result.BenchmarkProfile)
-	b.WriteString("- Single-GPU compute score from bee-gpu-burn cuBLASLt when available.\n")
-	b.WriteString("- Thermal and power limitations inferred from NVIDIA clock event reason counters and sustained telemetry.\n")
-	b.WriteString("- `result.json` is the canonical machine-readable source for this benchmark run.\n\n")
+	// ── Platform Scalability ──────────────────────────────────────────────────
+	if len(result.PerformanceRampSteps) > 0 {
+		b.WriteString("## Platform Scalability (Performance Ramp)\n\n")
+		fmt.Fprintf(&b, "**Platform power score:** %.1f%%  \n\n", result.PlatformPowerScore)
+		b.WriteString("| k GPUs | GPU Indices | Total Synthetic TOPS | Scalability |\n")
+		b.WriteString("|--------|-------------|----------------------|-------------|\n")
+		for _, step := range result.PerformanceRampSteps {
+			fmt.Fprintf(&b, "| %d | %s | %.2f | %.1f%% |\n",
+				step.StepIndex, joinIndexList(step.GPUIndices), step.TotalSyntheticTOPS, step.ScalabilityPct)
+		}
+		b.WriteString("\n")
+	}
 
 	// ── Raw files ─────────────────────────────────────────────────────────────
 	b.WriteString("## Raw Files\n\n")
 	b.WriteString("- `result.json`\n- `report.md`\n- `summary.txt`\n- `verbose.log`\n")
-	b.WriteString("- `gpu-*-baseline-metrics.csv/html/term.txt`\n")
-	b.WriteString("- `gpu-*-warmup.log`\n")
-	b.WriteString("- `gpu-*-steady.log`\n")
-	b.WriteString("- `gpu-*-steady-metrics.csv/html/term.txt`\n")
-	b.WriteString("- `gpu-*-cooldown-metrics.csv/html/term.txt`\n")
+	b.WriteString("- `gpu-metrics.csv`\n- `gpu-metrics.html`\n- `gpu-burn.log`\n")
 	if result.Interconnect != nil {
 		b.WriteString("- `nccl-all-reduce.log`\n")
 	}
 	return b.String()
 }
 
-// loadBenchmarkReportCharts loads only steady-state terminal charts (baseline and
-// cooldown charts are not useful for human review).
-func loadBenchmarkReportCharts(runDir string, gpuIndices []int) []benchmarkReportChart {
-	var charts []benchmarkReportChart
-	for _, idx := range gpuIndices {
-		path := filepath.Join(runDir, fmt.Sprintf("gpu-%d-steady-metrics-term.txt", idx))
-		raw, err := os.ReadFile(path)
-		if err != nil || len(raw) == 0 {
-			continue
-		}
-		charts = append(charts, benchmarkReportChart{
-			Title:   fmt.Sprintf("GPU %d — Steady State", idx),
-			Content: string(raw),
-		})
-	}
-	return charts
-}
-
-func stripANSIEscapeSequences(raw string) string {
-	return ansiEscapePattern.ReplaceAllString(raw, "")
-}
-
 // formatThrottleLine renders throttle counters as human-readable percentages of
 // the steady-state window.  Only non-zero counters are shown.  When the steady
 // duration is unknown (0), raw seconds are shown instead.
@@ -323,6 +470,7 @@ func formatThrottleLine(t BenchmarkThrottleCounters, steadyDurationSec float64)
 func renderBenchmarkSummary(result NvidiaBenchmarkResult) string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "run_at_utc=%s\n", result.GeneratedAt.Format(time.RFC3339))
+	fmt.Fprintf(&b, "benchmark_version=%s\n", result.BenchmarkVersion)
 	fmt.Fprintf(&b, "benchmark_profile=%s\n", result.BenchmarkProfile)
 	fmt.Fprintf(&b, "overall_status=%s\n", result.OverallStatus)
 	fmt.Fprintf(&b, "gpu_count=%d\n", len(result.GPUs))

audit/internal/platform/benchmark_test.go

@@ -16,17 +16,17 @@ func TestResolveBenchmarkProfile(t *testing.T) {
 		{
 			name:    "default",
 			profile: "",
-			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStandard, BaselineSec: 15, WarmupSec: 120, SteadySec: 480, NCCLSec: 180, CooldownSec: 120},
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStandard, BaselineSec: 15, WarmupSec: 45, SteadySec: 480, NCCLSec: 180, CooldownSec: 0},
 		},
 		{
 			name:    "stability",
 			profile: "stability",
-			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStability, BaselineSec: 30, WarmupSec: 300, SteadySec: 3600, NCCLSec: 300, CooldownSec: 300},
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStability, BaselineSec: 30, WarmupSec: 120, SteadySec: 3600, NCCLSec: 300, CooldownSec: 0},
 		},
 		{
 			name:    "overnight",
 			profile: "overnight",
-			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileOvernight, BaselineSec: 60, WarmupSec: 600, SteadySec: 27000, NCCLSec: 600, CooldownSec: 300},
+			want:    benchmarkProfileSpec{Name: NvidiaBenchmarkProfileOvernight, BaselineSec: 60, WarmupSec: 180, SteadySec: 27000, NCCLSec: 600, CooldownSec: 0},
 		},
 	}
 
@@ -41,6 +41,129 @@ func TestResolveBenchmarkProfile(t *testing.T) {
 	}
 }
 
+func TestBuildBenchmarkSteadyPlanStandard(t *testing.T) {
+	t.Parallel()
+
+	labels, phases, basePhaseSec, mixedPhaseSec := buildBenchmarkSteadyPlan(
+		benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStandard, SteadySec: 480},
+		benchmarkPrecisionPhases,
+		func(label string) string { return label },
+	)
+	if len(labels) != 5 || len(phases) != 5 {
+		t.Fatalf("labels=%d phases=%d want 5", len(labels), len(phases))
+	}
+	if basePhaseSec != 60 {
+		t.Fatalf("basePhaseSec=%d want 60", basePhaseSec)
+	}
+	if mixedPhaseSec != 300 {
+		t.Fatalf("mixedPhaseSec=%d want 300", mixedPhaseSec)
+	}
+	if phases[len(phases)-1].PlanLabel != "mixed" || phases[len(phases)-1].DurationSec != 300 {
+		t.Fatalf("mixed phase=%+v want duration 300", phases[len(phases)-1])
+	}
+	if benchmarkPlanDurationsCSV(phases) != "60,60,60,60,300" {
+		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
+	}
+}
+
+func TestBuildBenchmarkSteadyPlanStability(t *testing.T) {
+	t.Parallel()
+
+	_, phases, basePhaseSec, mixedPhaseSec := buildBenchmarkSteadyPlan(
+		benchmarkProfileSpec{Name: NvidiaBenchmarkProfileStability, SteadySec: 3600},
+		benchmarkPrecisionPhases,
+		func(label string) string { return label },
+	)
+	if basePhaseSec != 300 {
+		t.Fatalf("basePhaseSec=%d want 300", basePhaseSec)
+	}
+	if mixedPhaseSec != 3600 {
+		t.Fatalf("mixedPhaseSec=%d want 3600", mixedPhaseSec)
+	}
+	if benchmarkPlanDurationsCSV(phases) != "300,300,300,300,3600" {
+		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
+	}
+}
+
+func TestBuildBenchmarkSteadyPlanOvernight(t *testing.T) {
+	t.Parallel()
+
+	_, phases, basePhaseSec, mixedPhaseSec := buildBenchmarkSteadyPlan(
+		benchmarkProfileSpec{Name: NvidiaBenchmarkProfileOvernight, SteadySec: 27000},
+		benchmarkPrecisionPhases,
+		func(label string) string { return label },
+	)
+	if basePhaseSec != 3600 {
+		t.Fatalf("basePhaseSec=%d want 3600", basePhaseSec)
+	}
+	if mixedPhaseSec != 14400 {
+		t.Fatalf("mixedPhaseSec=%d want 14400", mixedPhaseSec)
+	}
+	if benchmarkPlanDurationsCSV(phases) != "3600,3600,3600,3600,14400" {
+		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
+	}
+}
+
+func TestSplitBenchmarkRowsByPlannedPhaseUsesPhaseDurations(t *testing.T) {
+	t.Parallel()
+
+	phases := []benchmarkPlannedPhase{
+		{PlanLabel: "fp8", MetricStage: "fp8", DurationSec: 10},
+		{PlanLabel: "fp16", MetricStage: "fp16", DurationSec: 10},
+		{PlanLabel: "mixed", MetricStage: "mixed", DurationSec: 50},
+	}
+	rows := []GPUMetricRow{
+		{ElapsedSec: 5},
+		{ElapsedSec: 15},
+		{ElapsedSec: 25},
+		{ElapsedSec: 65},
+	}
+	got := splitBenchmarkRowsByPlannedPhase(rows, phases)
+	if len(got["fp8"]) != 1 {
+		t.Fatalf("fp8 rows=%d want 1", len(got["fp8"]))
+	}
+	if len(got["fp16"]) != 1 {
+		t.Fatalf("fp16 rows=%d want 1", len(got["fp16"]))
+	}
+	if len(got["mixed"]) != 2 {
+		t.Fatalf("mixed rows=%d want 2", len(got["mixed"]))
+	}
+}
+
+func TestBenchmarkSupportedPrecisionsSkipsFP4BeforeBlackwell(t *testing.T) {
+	t.Parallel()
+
+	if got := benchmarkSupportedPrecisions("9.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32" {
+		t.Fatalf("supported=%v", got)
+	}
+	if got := benchmarkSupportedPrecisions("10.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32" {
+		t.Fatalf("supported=%v", got)
+	}
+}
+
+func TestBenchmarkPlannedPhaseStatus(t *testing.T) {
+	t.Parallel()
+
+	cases := []struct {
+		name       string
+		raw        string
+		wantStatus string
+	}{
+		{name: "ok", raw: "status=OK\n", wantStatus: "OK"},
+		{name: "failed", raw: "phase_error=fp16\n", wantStatus: "FAILED"},
+		{name: "unsupported", raw: "cublasLt_profiles=unsupported\nphase_error=fp4\n", wantStatus: "UNSUPPORTED"},
+	}
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			got, _ := benchmarkPlannedPhaseStatus([]byte(tc.raw))
+			if got != tc.wantStatus {
+				t.Fatalf("status=%q want %q", got, tc.wantStatus)
+			}
+		})
+	}
+}
+
 func TestNormalizeNvidiaBenchmarkOptionsPreservesRunNCCLChoice(t *testing.T) {
 	t.Parallel()
 
@@ -65,8 +188,10 @@ func TestParseBenchmarkBurnLog(t *testing.T) {
 		"[gpu 0] compute_capability=9.0",
 		"[gpu 0] backend=cublasLt",
 		"[gpu 0] duration_s=10",
+		"[gpu 0] int8_tensor[0]=READY dim=16384x16384x8192 block=128 stream=0",
 		"[gpu 0] fp16_tensor[0]=READY dim=4096x4096x4096 block=128 stream=0",
 		"[gpu 0] fp8_e4m3[0]=READY dim=8192x8192x4096 block=128 stream=0",
+		"[gpu 0] int8_tensor_iterations=80",
 		"[gpu 0] fp16_tensor_iterations=200",
 		"[gpu 0] fp8_e4m3_iterations=50",
 		"[gpu 0] status=OK",
@@ -79,15 +204,24 @@ func TestParseBenchmarkBurnLog(t *testing.T) {
 	if got.ComputeCapability != "9.0" {
 		t.Fatalf("compute capability=%q want 9.0", got.ComputeCapability)
 	}
-	if len(got.Profiles) != 2 {
-		t.Fatalf("profiles=%d want 2", len(got.Profiles))
+	if len(got.Profiles) != 3 {
+		t.Fatalf("profiles=%d want 3", len(got.Profiles))
 	}
 	if got.Profiles[0].TeraOpsPerSec <= 0 {
 		t.Fatalf("profile[0] teraops=%f want >0", got.Profiles[0].TeraOpsPerSec)
 	}
+	if got.Profiles[0].Category != "fp16_bf16" {
+		t.Fatalf("profile[0] category=%q want fp16_bf16", got.Profiles[0].Category)
+	}
 	if got.Profiles[1].Category != "fp8" {
 		t.Fatalf("profile[1] category=%q want fp8", got.Profiles[1].Category)
 	}
+	if got.Profiles[2].Category != "int8" {
+		t.Fatalf("profile[2] category=%q want int8", got.Profiles[2].Category)
+	}
+	if got.Profiles[2].Weight != 0.25 {
+		t.Fatalf("profile[2] weight=%f want 0.25", got.Profiles[2].Weight)
+	}
 }
 
 func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
@@ -131,6 +265,13 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 				DegradationReasons: []string{"power_capped"},
 			},
 		},
+		Cooling: &BenchmarkCoolingSummary{
+			Available:             true,
+			AvgFanRPM:             9200,
+			FanDutyCycleAvailable: true,
+			AvgFanDutyCyclePct:    47.5,
+			P95FanDutyCyclePct:    62.0,
+		},
 	}
 
 	report := renderBenchmarkReport(result)
@@ -140,6 +281,9 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 		"1176.00",
 		"fp16_tensor",
 		"700.00",
+		"Cooling",
+		"Average fan duty cycle",
+		"47.5%",
 	} {
 		if !strings.Contains(report, needle) {
 			t.Fatalf("report missing %q\n%s", needle, report)
@@ -147,34 +291,113 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 	}
 }
 
-func TestRenderBenchmarkReportIncludesTerminalChartsWithoutANSI(t *testing.T) {
+func TestRenderBenchmarkReportListsUnifiedArtifacts(t *testing.T) {
 	t.Parallel()
 
-	report := renderBenchmarkReportWithCharts(NvidiaBenchmarkResult{
+	report := renderBenchmarkReport(NvidiaBenchmarkResult{
 		BenchmarkProfile:   NvidiaBenchmarkProfileStandard,
 		OverallStatus:      "OK",
 		SelectedGPUIndices: []int{0},
 		Normalization: BenchmarkNormalization{
 			Status: "full",
 		},
-	}, []benchmarkReportChart{
-		{
-			Title:   "GPU 0 Steady State",
-			Content: "\x1b[31mGPU 0 chart\x1b[0m\n 42┤───",
-		},
 	})
 
 	for _, needle := range []string{
-		"Steady-State Charts",
-		"GPU 0 Steady State",
-		"GPU 0 chart",
-		"42┤───",
+		"gpu-metrics.csv",
+		"gpu-metrics.html",
+		"gpu-burn.log",
 	} {
 		if !strings.Contains(report, needle) {
 			t.Fatalf("report missing %q\n%s", needle, report)
 		}
 	}
-	if strings.Contains(report, "\x1b[31m") {
-		t.Fatalf("report should not contain ANSI escapes\n%s", report)
+}
+
+func TestScoreBenchmarkGPUIgnoresDisabledPrecisions(t *testing.T) {
+	t.Parallel()
+
+	score := scoreBenchmarkGPUResult(BenchmarkGPUResult{
+		PrecisionSteady: []BenchmarkPrecisionSteadyPhase{
+			{Precision: "fp16", WeightedTeraOpsPerSec: 100},
+			{Precision: "fp64", WeightedTeraOpsPerSec: 999},
+			{Precision: "fp4", WeightedTeraOpsPerSec: 999},
+		},
+		PrecisionResults: []BenchmarkPrecisionResult{
+			{Category: "fp32_tf32", Supported: true, WeightedTeraOpsPerSec: 50},
+			{Category: "fp64", Supported: true, WeightedTeraOpsPerSec: 999},
+			{Category: "fp4", Supported: true, WeightedTeraOpsPerSec: 999},
+		},
+	})
+
+	if score.SyntheticScore != 100 {
+		t.Fatalf("SyntheticScore=%f want 100", score.SyntheticScore)
+	}
+	if score.MixedScore != 50 {
+		t.Fatalf("MixedScore=%f want 50", score.MixedScore)
+	}
+}
+
+func TestEnrichGPUInfoWithMaxClocks(t *testing.T) {
+	t.Parallel()
+
+	nvsmiQ := []byte(`
+GPU 00000000:4E:00.0
+    Product Name                          : NVIDIA RTX PRO 6000 Blackwell Server Edition
+    Clocks
+        Graphics                          : 2422 MHz
+        Memory                            : 12481 MHz
+    Max Clocks
+        Graphics                          : 2430 MHz
+        SM                                : 2430 MHz
+        Memory                            : 12481 MHz
+        Video                             : 2107 MHz
+
+GPU 00000000:4F:00.0
+    Product Name                          : NVIDIA RTX PRO 6000 Blackwell Server Edition
+    Max Clocks
+        Graphics                          : 2430 MHz
+        Memory                            : 12481 MHz
+`)
+
+	infoByIndex := map[int]benchmarkGPUInfo{
+		0: {Index: 0, BusID: "00000000:4E:00.0"},
+		1: {Index: 1, BusID: "00000000:4F:00.0"},
+	}
+
+	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+
+	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("GPU 0 MaxGraphicsClockMHz = %v, want 2430", infoByIndex[0].MaxGraphicsClockMHz)
+	}
+	if infoByIndex[0].MaxMemoryClockMHz != 12481 {
+		t.Errorf("GPU 0 MaxMemoryClockMHz = %v, want 12481", infoByIndex[0].MaxMemoryClockMHz)
+	}
+	if infoByIndex[1].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("GPU 1 MaxGraphicsClockMHz = %v, want 2430", infoByIndex[1].MaxGraphicsClockMHz)
+	}
+	if infoByIndex[1].MaxMemoryClockMHz != 12481 {
+		t.Errorf("GPU 1 MaxMemoryClockMHz = %v, want 12481", infoByIndex[1].MaxMemoryClockMHz)
+	}
+}
+
+func TestEnrichGPUInfoWithMaxClocksSkipsPopulated(t *testing.T) {
+	t.Parallel()
+
+	nvsmiQ := []byte(`
+GPU 00000000:4E:00.0
+    Max Clocks
+        Graphics                          : 9999 MHz
+        Memory                            : 9999 MHz
+`)
+	// Already populated — must not be overwritten.
+	infoByIndex := map[int]benchmarkGPUInfo{
+		0: {Index: 0, BusID: "00000000:4E:00.0", MaxGraphicsClockMHz: 2430, MaxMemoryClockMHz: 12481},
+	}
+
+	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+
+	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
+		t.Errorf("expected existing value to be preserved, got %v", infoByIndex[0].MaxGraphicsClockMHz)
 	}
 }

audit/internal/platform/benchmark_types.go

@@ -2,6 +2,41 @@ package platform
 
 import "time"
 
+// BenchmarkHostConfig holds static CPU and memory configuration captured at
+// benchmark start. Useful for correlating results across runs on different hardware.
+type BenchmarkHostConfig struct {
+	CPUModel    string  `json:"cpu_model,omitempty"`
+	CPUSockets  int     `json:"cpu_sockets,omitempty"`
+	CPUCores    int     `json:"cpu_cores,omitempty"`
+	CPUThreads  int     `json:"cpu_threads,omitempty"`
+	MemTotalGiB float64 `json:"mem_total_gib,omitempty"`
+}
+
+// BenchmarkCPULoad summarises host CPU utilisation sampled during the GPU
+// steady-state phase. High or unstable CPU load during a GPU benchmark may
+// indicate a competing workload or a CPU-bound driver bottleneck.
+type BenchmarkCPULoad struct {
+	AvgPct  float64 `json:"avg_pct"`
+	MaxPct  float64 `json:"max_pct"`
+	P95Pct  float64 `json:"p95_pct"`
+	Samples int     `json:"samples"`
+	// Status is "ok", "high", or "unstable".
+	Status string `json:"status"`
+	Note   string `json:"note,omitempty"`
+}
+
+// BenchmarkCoolingSummary captures fan telemetry averaged across the full
+// benchmark run.
+type BenchmarkCoolingSummary struct {
+	Available             bool     `json:"available"`
+	AvgFanRPM             float64  `json:"avg_fan_rpm,omitempty"`
+	FanDutyCycleAvailable bool     `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool     `json:"fan_duty_cycle_estimated,omitempty"`
+	AvgFanDutyCyclePct    float64  `json:"avg_fan_duty_cycle_pct,omitempty"`
+	P95FanDutyCyclePct    float64  `json:"p95_fan_duty_cycle_pct,omitempty"`
+	Notes                 []string `json:"notes,omitempty"`
+}
+
 const (
 	NvidiaBenchmarkProfileStandard  = "standard"
 	NvidiaBenchmarkProfileStability = "stability"
@@ -14,25 +49,39 @@ type NvidiaBenchmarkOptions struct {
 	GPUIndices        []int
 	ExcludeGPUIndices []int
 	RunNCCL           bool
-	ParallelGPUs      bool // run all selected GPUs simultaneously instead of sequentially
+	ParallelGPUs      bool   // run all selected GPUs simultaneously instead of sequentially
+	RampStep          int    // 1-based step index within a ramp-up run (0 = not a ramp-up)
+	RampTotal         int    // total number of ramp-up steps in this run
+	RampRunID         string // shared identifier across all steps of the same ramp-up run
 }
 
-
 type NvidiaBenchmarkResult struct {
-	BenchmarkVersion   string                       `json:"benchmark_version"`
-	GeneratedAt        time.Time                    `json:"generated_at"`
-	Hostname           string                       `json:"hostname,omitempty"`
-	ServerModel        string                       `json:"server_model,omitempty"`
-	BenchmarkProfile   string                       `json:"benchmark_profile"`
-	ParallelGPUs       bool                         `json:"parallel_gpus,omitempty"`
-	OverallStatus      string                       `json:"overall_status"`
-	SelectedGPUIndices []int                        `json:"selected_gpu_indices"`
-	Findings           []string                     `json:"findings,omitempty"`
-	Warnings           []string                     `json:"warnings,omitempty"`
-	Normalization      BenchmarkNormalization       `json:"normalization"`
-	GPUs               []BenchmarkGPUResult         `json:"gpus"`
-	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
-	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`
+	BenchmarkVersion string    `json:"benchmark_version"`
+	GeneratedAt      time.Time `json:"generated_at"`
+	Hostname         string    `json:"hostname,omitempty"`
+	ServerModel      string    `json:"server_model,omitempty"`
+	BenchmarkProfile string    `json:"benchmark_profile"`
+	ParallelGPUs     bool      `json:"parallel_gpus,omitempty"`
+	RampStep         int       `json:"ramp_step,omitempty"`
+	RampTotal        int       `json:"ramp_total,omitempty"`
+	RampRunID        string    `json:"ramp_run_id,omitempty"`
+	ScalabilityScore float64   `json:"scalability_score,omitempty"`
+	// PlatformPowerScore is the mean compute scalability across ramp steps 2..N.
+	// 100% = each added GPU contributes exactly its single-card throughput.
+	// < 100% = throughput loss due to thermal throttle, power limits, or contention.
+	PlatformPowerScore   float64                      `json:"platform_power_score,omitempty"`
+	PerformanceRampSteps []NvidiaPerformanceRampStep  `json:"performance_ramp_steps,omitempty"`
+	OverallStatus        string                       `json:"overall_status"`
+	SelectedGPUIndices   []int                        `json:"selected_gpu_indices"`
+	Findings             []string                     `json:"findings,omitempty"`
+	Warnings             []string                     `json:"warnings,omitempty"`
+	Normalization        BenchmarkNormalization       `json:"normalization"`
+	HostConfig           *BenchmarkHostConfig         `json:"host_config,omitempty"`
+	CPULoad              *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
+	Cooling              *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
+	GPUs                 []BenchmarkGPUResult         `json:"gpus"`
+	Interconnect         *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
+	ServerPower          *BenchmarkServerPower        `json:"server_power,omitempty"`
 }
 
 type BenchmarkNormalization struct {
@@ -52,30 +101,51 @@ type BenchmarkNormalizationGPU struct {
 }
 
 type BenchmarkGPUResult struct {
-	Index                  int                        `json:"index"`
-	UUID                   string                     `json:"uuid,omitempty"`
-	Name                   string                     `json:"name,omitempty"`
-	BusID                  string                     `json:"bus_id,omitempty"`
-	VBIOS                  string                     `json:"vbios,omitempty"`
-	ComputeCapability      string                     `json:"compute_capability,omitempty"`
-	Backend                string                     `json:"backend,omitempty"`
-	Status                 string                     `json:"status"`
-	PowerLimitW            float64                    `json:"power_limit_w,omitempty"`
-	MultiprocessorCount    int                        `json:"multiprocessor_count,omitempty"`
-	DefaultPowerLimitW     float64                    `json:"default_power_limit_w,omitempty"`
-	MaxGraphicsClockMHz    float64                    `json:"max_graphics_clock_mhz,omitempty"`
-	BaseGraphicsClockMHz   float64                    `json:"base_graphics_clock_mhz,omitempty"`
-	MaxMemoryClockMHz      float64                    `json:"max_memory_clock_mhz,omitempty"`
-	LockedGraphicsClockMHz float64                    `json:"locked_graphics_clock_mhz,omitempty"`
-	LockedMemoryClockMHz   float64                    `json:"locked_memory_clock_mhz,omitempty"`
-	Baseline               BenchmarkTelemetrySummary  `json:"baseline"`
-	Steady                 BenchmarkTelemetrySummary  `json:"steady"`
-	Cooldown               BenchmarkTelemetrySummary  `json:"cooldown"`
-	Throttle               BenchmarkThrottleCounters  `json:"throttle_counters"`
-	PrecisionResults       []BenchmarkPrecisionResult `json:"precision_results,omitempty"`
-	Scores                 BenchmarkScorecard         `json:"scores"`
-	DegradationReasons     []string                   `json:"degradation_reasons,omitempty"`
-	Notes                  []string                   `json:"notes,omitempty"`
+	Index               int     `json:"index"`
+	UUID                string  `json:"uuid,omitempty"`
+	Name                string  `json:"name,omitempty"`
+	BusID               string  `json:"bus_id,omitempty"`
+	VBIOS               string  `json:"vbios,omitempty"`
+	ComputeCapability   string  `json:"compute_capability,omitempty"`
+	Backend             string  `json:"backend,omitempty"`
+	Status              string  `json:"status"`
+	PowerLimitW         float64 `json:"power_limit_w,omitempty"`
+	PowerLimitDerated   bool    `json:"power_limit_derated,omitempty"`
+	MultiprocessorCount int     `json:"multiprocessor_count,omitempty"`
+	DefaultPowerLimitW  float64 `json:"default_power_limit_w,omitempty"`
+	// ShutdownTempC is the hardware thermal shutdown threshold for this GPU,
+	// sourced from nvidia-smi -q ("GPU Shutdown Temp"). Fallback: 90°C.
+	ShutdownTempC float64 `json:"shutdown_temp_c,omitempty"`
+	// SlowdownTempC is the software throttle onset threshold ("GPU Slowdown Temp").
+	// Fallback: 80°C.
+	SlowdownTempC float64 `json:"slowdown_temp_c,omitempty"`
+	// CalibratedPeakPowerW is the p95 power measured during a short
+	// dcgmi targeted_power calibration run before the main benchmark.
+	// Used as the reference denominator for PowerSustainScore instead of
+	// the hardware default limit, which bee-gpu-burn cannot reach.
+	CalibratedPeakPowerW   float64                         `json:"calibrated_peak_power_w,omitempty"`
+	CalibratedPeakTempC    float64                         `json:"calibrated_peak_temp_c,omitempty"`
+	PowerCalibrationTries  int                             `json:"power_calibration_tries,omitempty"`
+	MaxGraphicsClockMHz    float64                         `json:"max_graphics_clock_mhz,omitempty"`
+	BaseGraphicsClockMHz   float64                         `json:"base_graphics_clock_mhz,omitempty"`
+	MaxMemoryClockMHz      float64                         `json:"max_memory_clock_mhz,omitempty"`
+	LockedGraphicsClockMHz float64                         `json:"locked_graphics_clock_mhz,omitempty"`
+	LockedMemoryClockMHz   float64                         `json:"locked_memory_clock_mhz,omitempty"`
+	Baseline               BenchmarkTelemetrySummary       `json:"baseline"`
+	Steady                 BenchmarkTelemetrySummary       `json:"steady"`
+	PrecisionSteady        []BenchmarkPrecisionSteadyPhase `json:"precision_steady,omitempty"`
+	PrecisionFailures      []string                        `json:"precision_failures,omitempty"`
+	Cooldown               BenchmarkTelemetrySummary       `json:"cooldown"`
+	Throttle               BenchmarkThrottleCounters       `json:"throttle_counters"`
+	// ECC error delta accumulated over the full benchmark (all phases combined).
+	ECC                BenchmarkECCCounters       `json:"ecc,omitempty"`
+	PrecisionResults   []BenchmarkPrecisionResult `json:"precision_results,omitempty"`
+	Scores             BenchmarkScorecard         `json:"scores"`
+	DegradationReasons []string                   `json:"degradation_reasons,omitempty"`
+	Notes              []string                   `json:"notes,omitempty"`
+	// CoolingWarning is non-empty when a thermal throttle event occurred with
+	// a clock drop ≥20% while server fans were not at 100% duty cycle.
+	CoolingWarning string `json:"cooling_warning,omitempty"`
 }
 
 type BenchmarkTelemetrySummary struct {
@@ -105,6 +175,18 @@ type BenchmarkThrottleCounters struct {
 	HWPowerBrakeSlowdownUS uint64 `json:"hw_power_brake_slowdown_us"`
 }
 
+// BenchmarkECCCounters holds ECC error counts sampled at a point in time.
+// Corrected = single-bit errors fixed by ECC (DRAM degradation).
+// Uncorrected = double-bit errors that could not be corrected (serious fault).
+// Both are volatile (since last driver reset), not persistent.
+type BenchmarkECCCounters struct {
+	Corrected   uint64 `json:"corrected"`
+	Uncorrected uint64 `json:"uncorrected"`
+}
+
+func (e BenchmarkECCCounters) Total() uint64 { return e.Corrected + e.Uncorrected }
+func (e BenchmarkECCCounters) IsZero() bool  { return e.Corrected == 0 && e.Uncorrected == 0 }
+
 type BenchmarkPrecisionResult struct {
 	Name          string  `json:"name"`
 	Category      string  `json:"category"`
@@ -115,19 +197,52 @@ type BenchmarkPrecisionResult struct {
 	K             uint64  `json:"k,omitempty"`
 	Iterations    uint64  `json:"iterations,omitempty"`
 	TeraOpsPerSec float64 `json:"teraops_per_sec,omitempty"`
-	Notes         string  `json:"notes,omitempty"`
+	// Weight is the fp32-equivalence factor for this precision category.
+	// fp32 = 1.0 (baseline), fp64 = 2.0, fp16 = 0.5, int8/fp8 = 0.25, fp4 = 0.125.
+	// WeightedTOPS = TeraOpsPerSec * Weight gives fp32-equivalent throughput.
+	Weight                float64 `json:"weight,omitempty"`
+	WeightedTeraOpsPerSec float64 `json:"weighted_teraops_per_sec,omitempty"`
+	Notes                 string  `json:"notes,omitempty"`
 }
 
 type BenchmarkScorecard struct {
-	ComputeScore        float64 `json:"compute_score"`
+	ComputeScore float64 `json:"compute_score"`
+	// SyntheticScore is the sum of fp32-equivalent TOPS from per-precision
+	// steady phases (each precision ran alone, full GPU dedicated).
+	SyntheticScore float64 `json:"synthetic_score,omitempty"`
+	// MixedScore is the sum of fp32-equivalent TOPS from the combined phase
+	// (all precisions competing simultaneously — closer to real workloads).
+	MixedScore float64 `json:"mixed_score,omitempty"`
+	// MixedEfficiency = MixedScore / SyntheticScore. Measures how well the GPU
+	// sustains throughput under concurrent mixed-precision load.
+	MixedEfficiency     float64 `json:"mixed_efficiency,omitempty"`
 	PowerSustainScore   float64 `json:"power_sustain_score"`
 	ThermalSustainScore float64 `json:"thermal_sustain_score"`
-	StabilityScore      float64 `json:"stability_score"`
-	InterconnectScore   float64 `json:"interconnect_score"`
-	CompositeScore      float64 `json:"composite_score"`
+	// StabilityScore: fraction of steady-state time the GPU spent throttling
+	// (thermal + power cap combined). 0% throttle = 100; 100% throttle = 0.
+	StabilityScore float64 `json:"stability_score"`
+
+	// Throttle breakdown — percentage of steady-state time in each throttle type.
+	// Used for diagnosis: tells WHY the GPU throttled, not just whether it did.
+	ThermalThrottlePct   float64 `json:"thermal_throttle_pct"`   // HW+SW thermal slowdown
+	PowerCapThrottlePct  float64 `json:"power_cap_throttle_pct"` // SW power cap
+	SyncBoostThrottlePct float64 `json:"sync_boost_throttle_pct,omitempty"`
+
+	// Temperature headroom: distance to the 100°C destruction threshold.
+	// TempHeadroomC = 100 - P95TempC. < 20°C = warning; < 10°C = critical.
+	// Independent of throttle — a GPU at 86°C without throttle is still in the red zone.
+	TempHeadroomC float64 `json:"temp_headroom_c"`
+
+	InterconnectScore float64 `json:"interconnect_score"`
+	// ServerQualityScore (0–100) reflects server infrastructure quality independent
+	// of GPU model. Combines throttle time, power variance, and temp variance.
+	// Use this to compare servers with the same GPU, or to flag a bad server
+	// that throttles an otherwise fast GPU.
+	ServerQualityScore float64 `json:"server_quality_score"`
+	// CompositeScore is the raw compute score (TOPS, fp32-equivalent).
+	// A throttling GPU will score lower here automatically — no quality multiplier.
+	CompositeScore float64 `json:"composite_score"`
 	// TOPSPerSMPerGHz is compute efficiency independent of clock speed and SM count.
-	// Comparable across throttle levels and GPU generations. Low value at normal
-	// clocks indicates silicon degradation.
 	TOPSPerSMPerGHz float64 `json:"tops_per_sm_per_ghz,omitempty"`
 }
 
@@ -145,6 +260,22 @@ type BenchmarkServerPower struct {
 	Notes           []string `json:"notes,omitempty"`
 }
 
+// BenchmarkPrecisionSteadyPhase holds per-precision-category telemetry collected
+// during a dedicated single-precision steady window.  Because only one kernel
+// type runs at a time the PowerCVPct here is a genuine stability signal.
+type BenchmarkPrecisionSteadyPhase struct {
+	Precision             string                    `json:"precision"` // e.g. "fp8", "fp16", "fp32"
+	Status                string                    `json:"status,omitempty"`
+	Steady                BenchmarkTelemetrySummary `json:"steady"`
+	TeraOpsPerSec         float64                   `json:"teraops_per_sec,omitempty"`
+	WeightedTeraOpsPerSec float64                   `json:"weighted_teraops_per_sec,omitempty"`
+	// ECC errors accumulated during this precision phase only.
+	// Non-zero corrected = stress-induced DRAM errors for this kernel type.
+	// Any uncorrected = serious fault triggered by this precision workload.
+	ECC   BenchmarkECCCounters `json:"ecc,omitempty"`
+	Notes string               `json:"notes,omitempty"`
+}
+
 type BenchmarkInterconnectResult struct {
 	Status             string   `json:"status"`
 	Attempted          bool     `json:"attempted"`
@@ -156,3 +287,78 @@ type BenchmarkInterconnectResult struct {
 	MaxBusBWGBps       float64  `json:"max_busbw_gbps,omitempty"`
 	Notes              []string `json:"notes,omitempty"`
 }
+
+type NvidiaPowerBenchResult struct {
+	BenchmarkVersion     string                 `json:"benchmark_version"`
+	GeneratedAt          time.Time              `json:"generated_at"`
+	Hostname             string                 `json:"hostname,omitempty"`
+	ServerModel          string                 `json:"server_model,omitempty"`
+	BenchmarkProfile     string                 `json:"benchmark_profile"`
+	SelectedGPUIndices   []int                  `json:"selected_gpu_indices"`
+	RecommendedSlotOrder []int                  `json:"recommended_slot_order,omitempty"`
+	RampSteps            []NvidiaPowerBenchStep `json:"ramp_steps,omitempty"`
+	OverallStatus        string                 `json:"overall_status"`
+	// PlatformMaxTDPW is the sum of per-GPU stable power limits found during the
+	// cumulative thermal ramp. Represents the actual sustained power budget of
+	// this server under full GPU load. Use for rack power planning.
+	PlatformMaxTDPW float64 `json:"platform_max_tdp_w"`
+	// ServerPower captures IPMI server power delta (idle→loaded) measured in
+	// parallel with the thermal ramp. Use to compare GPU-reported TDP against
+	// actual wall-power draw as seen by the server's power supply.
+	ServerPower *BenchmarkServerPower `json:"server_power,omitempty"`
+	Findings    []string              `json:"findings,omitempty"`
+	GPUs        []NvidiaPowerBenchGPU `json:"gpus"`
+}
+
+type NvidiaPowerBenchGPU struct {
+	Index              int     `json:"index"`
+	Name               string  `json:"name,omitempty"`
+	BusID              string  `json:"bus_id,omitempty"`
+	DefaultPowerLimitW float64 `json:"default_power_limit_w,omitempty"`
+	// AppliedPowerLimitW is the stable limit found during single-card calibration.
+	AppliedPowerLimitW float64 `json:"applied_power_limit_w,omitempty"`
+	// StablePowerLimitW is the final fixed limit for this GPU after the
+	// cumulative thermal ramp. This is the limit at which the GPU operated
+	// stably with all other GPUs running simultaneously at their own limits.
+	// May be lower than AppliedPowerLimitW if multi-GPU thermal load required
+	// additional derating.
+	StablePowerLimitW   float64  `json:"stable_power_limit_w,omitempty"`
+	MaxObservedPowerW   float64  `json:"max_observed_power_w,omitempty"`
+	MaxObservedTempC    float64  `json:"max_observed_temp_c,omitempty"`
+	CalibrationAttempts int      `json:"calibration_attempts,omitempty"`
+	Derated             bool     `json:"derated,omitempty"`
+	Status              string   `json:"status"`
+	Notes               []string `json:"notes,omitempty"`
+	// CoolingWarning mirrors BenchmarkGPUResult.CoolingWarning for the power workflow.
+	CoolingWarning string `json:"cooling_warning,omitempty"`
+}
+
+type NvidiaPowerBenchStep struct {
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
+	// NewGPUIndex is the GPU whose stable limit was searched in this step.
+	NewGPUIndex int `json:"new_gpu_index"`
+	// NewGPUStableLimitW is the stable power limit found for the new GPU.
+	NewGPUStableLimitW  float64  `json:"new_gpu_stable_limit_w,omitempty"`
+	TotalObservedPowerW float64  `json:"total_observed_power_w,omitempty"`
+	AvgObservedPowerW   float64  `json:"avg_observed_power_w,omitempty"`
+	Derated             bool     `json:"derated,omitempty"`
+	Status              string   `json:"status"`
+	Notes               []string `json:"notes,omitempty"`
+}
+
+// NvidiaPerformanceRampStep holds per-step performance data for the
+// scalability ramp-up phase of the performance benchmark.
+type NvidiaPerformanceRampStep struct {
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
+	// TotalSyntheticTOPS is the sum of per-GPU SyntheticScore (fp32-equivalent
+	// TOPS from dedicated single-precision phases) across all GPUs in this step.
+	TotalSyntheticTOPS float64 `json:"total_synthetic_tops"`
+	TotalMixedTOPS     float64 `json:"total_mixed_tops,omitempty"`
+	// ScalabilityPct = TotalSyntheticTOPS / (k × best_single_gpu_tops) × 100.
+	// 100% = perfect linear scaling. < 100% = thermal/power/interconnect loss.
+	ScalabilityPct float64  `json:"scalability_pct"`
+	Status         string   `json:"status"`
+	Notes          []string `json:"notes,omitempty"`
+}

audit/internal/platform/gpu_metrics.go

@@ -13,14 +13,21 @@ import (
 
 // GPUMetricRow is one telemetry sample from nvidia-smi during a stress test.
 type GPUMetricRow struct {
-	ElapsedSec  float64 `json:"elapsed_sec"`
-	GPUIndex    int     `json:"index"`
-	TempC       float64 `json:"temp_c"`
-	UsagePct    float64 `json:"usage_pct"`
-	MemUsagePct float64 `json:"mem_usage_pct"`
-	PowerW      float64 `json:"power_w"`
-	ClockMHz    float64 `json:"clock_mhz"`
-	MemClockMHz float64 `json:"mem_clock_mhz"`
+	Stage                 string  `json:"stage,omitempty"`
+	StageStartSec         float64 `json:"stage_start_sec,omitempty"`
+	StageEndSec           float64 `json:"stage_end_sec,omitempty"`
+	ElapsedSec            float64 `json:"elapsed_sec"`
+	GPUIndex              int     `json:"index"`
+	TempC                 float64 `json:"temp_c"`
+	UsagePct              float64 `json:"usage_pct"`
+	MemUsagePct           float64 `json:"mem_usage_pct"`
+	PowerW                float64 `json:"power_w"`
+	ClockMHz              float64 `json:"clock_mhz"`
+	MemClockMHz           float64 `json:"mem_clock_mhz"`
+	FanAvgRPM             float64 `json:"fan_avg_rpm,omitempty"`
+	FanDutyCyclePct       float64 `json:"fan_duty_cycle_pct,omitempty"`
+	FanDutyCycleAvailable bool    `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool    `json:"fan_duty_cycle_estimated,omitempty"`
 }
 
 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
@@ -141,14 +148,28 @@ func sampleAMDGPUMetrics() ([]GPUMetricRow, error) {
 // WriteGPUMetricsCSV writes collected rows as a CSV file.
 func WriteGPUMetricsCSV(path string, rows []GPUMetricRow) error {
 	var b bytes.Buffer
-	b.WriteString("elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz\n")
+	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz,fan_avg_rpm,fan_duty_cycle_pct,fan_duty_cycle_available,fan_duty_cycle_estimated\n")
 	for _, r := range rows {
-		fmt.Fprintf(&b, "%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f\n",
-			r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz)
+		dutyAvail := 0
+		if r.FanDutyCycleAvailable {
+			dutyAvail = 1
+		}
+		dutyEstimated := 0
+		if r.FanDutyCycleEstimated {
+			dutyEstimated = 1
+		}
+		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%d,%d\n",
+			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz, r.FanAvgRPM, r.FanDutyCyclePct, dutyAvail, dutyEstimated)
 	}
 	return os.WriteFile(path, b.Bytes(), 0644)
 }
 
+type gpuMetricStageSpan struct {
+	Name  string
+	Start float64
+	End   float64
+}
+
 // WriteGPUMetricsHTML writes a standalone HTML file with one SVG chart per GPU.
 func WriteGPUMetricsHTML(path string, rows []GPUMetricRow) error {
 	// Group by GPU index preserving order.
@@ -163,9 +184,25 @@ func WriteGPUMetricsHTML(path string, rows []GPUMetricRow) error {
 		gpuMap[r.GPUIndex] = append(gpuMap[r.GPUIndex], r)
 	}
 
+	stageSpans := buildGPUMetricStageSpans(rows)
+	stageColorByName := make(map[string]string, len(stageSpans))
+	for i, span := range stageSpans {
+		stageColorByName[span.Name] = gpuMetricStagePalette[i%len(gpuMetricStagePalette)]
+	}
+
+	var legend strings.Builder
+	if len(stageSpans) > 0 {
+		legend.WriteString(`<div class="stage-legend">`)
+		for _, span := range stageSpans {
+			fmt.Fprintf(&legend, `<span class="stage-chip"><span class="stage-swatch" style="background:%s"></span>%s</span>`,
+				stageColorByName[span.Name], gpuHTMLEscape(span.Name))
+		}
+		legend.WriteString(`</div>`)
+	}
+
 	var svgs strings.Builder
 	for _, gpuIdx := range order {
-		svgs.WriteString(drawGPUChartSVG(gpuMap[gpuIdx], gpuIdx))
+		svgs.WriteString(drawGPUChartSVG(gpuMap[gpuIdx], gpuIdx, stageSpans, stageColorByName))
 		svgs.WriteString("\n")
 	}
 
@@ -175,21 +212,39 @@ func WriteGPUMetricsHTML(path string, rows []GPUMetricRow) error {
 <meta charset="utf-8">
 <title>GPU Stress Test Metrics</title>
 <style>
-body { font-family: sans-serif; background: #f0f0f0; margin: 0; padding: 20px; }
-h1 { text-align: center; color: #333; margin: 0 0 8px; }
-p  { text-align: center; color: #888; font-size: 13px; margin: 0 0 24px; }
+:root{--bg:#fff;--surface:#fff;--surface-2:#f9fafb;--border:rgba(34,36,38,.15);--border-lite:rgba(34,36,38,.1);--ink:rgba(0,0,0,.87);--muted:rgba(0,0,0,.6)}
+*{box-sizing:border-box}
+body{font:14px/1.5 Lato,"Helvetica Neue",Arial,Helvetica,sans-serif;background:var(--bg);color:var(--ink);margin:0}
+.page{padding:24px}
+.card{background:var(--surface);border:1px solid var(--border);border-radius:4px;box-shadow:0 1px 2px rgba(34,36,38,.15);overflow:hidden}
+.card-head{padding:11px 16px;background:var(--surface-2);border-bottom:1px solid var(--border);font-weight:700;font-size:13px}
+.card-body{padding:16px}
+h1{font-size:22px;margin:0 0 6px}
+p{color:var(--muted);font-size:13px;margin:0 0 16px}
+.stage-legend{display:flex;flex-wrap:wrap;gap:10px;margin:0 0 16px}
+.stage-chip{display:inline-flex;align-items:center;gap:8px;padding:4px 10px;border-radius:999px;background:var(--surface-2);border:1px solid var(--border-lite);font-size:12px}
+.stage-swatch{display:inline-block;width:12px;height:12px;border-radius:999px}
+.chart-block{margin-top:16px}
 </style>
 </head><body>
+<div class="page">
+<div class="card">
+<div class="card-head">GPU Stress Test Metrics</div>
+<div class="card-body">
 <h1>GPU Stress Test Metrics</h1>
 <p>Generated %s</p>
 %s
-</body></html>`, ts, svgs.String())
+<div class="chart-block">%s</div>
+</div>
+</div>
+</div>
+</body></html>`, ts, legend.String(), svgs.String())
 
 	return os.WriteFile(path, []byte(html), 0644)
 }
 
 // drawGPUChartSVG generates a self-contained SVG chart for one GPU.
-func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
+func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int, stageSpans []gpuMetricStageSpan, stageColorByName map[string]string) string {
 	// Layout
 	const W, H = 960, 520
 	const plotX1 = 120 // usage axis / chart left border
@@ -284,6 +339,23 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 	}
 	b.WriteString("</g>\n")
 
+	// Stage backgrounds
+	for _, span := range stageSpans {
+		x1 := xv(span.Start)
+		x2 := xv(span.End)
+		if x2 < x1 {
+			x1, x2 = x2, x1
+		}
+		if x2-x1 < 1 {
+			x2 = x1 + 1
+		}
+		color := stageColorByName[span.Name]
+		fmt.Fprintf(&b, `<rect x="%.1f" y="%d" width="%.1f" height="%d" fill="%s" fill-opacity="0.18"/>`+"\n",
+			x1, plotY1, x2-x1, PH, color)
+		fmt.Fprintf(&b, `<text x="%.1f" y="%d" font-family="sans-serif" font-size="10" fill="#444" text-anchor="middle">%s</text>`+"\n",
+			x1+(x2-x1)/2, plotY1+12, gpuHTMLEscape(span.Name))
+	}
+
 	// Chart border
 	fmt.Fprintf(&b, `<rect x="%d" y="%d" width="%d" height="%d"`+
 		` fill="none" stroke="#333" stroke-width="1"/>`+"\n",
@@ -382,224 +454,6 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 	return b.String()
 }
 
-const (
-	ansiRed    = "\033[31m"
-	ansiBlue   = "\033[34m"
-	ansiGreen  = "\033[32m"
-	ansiYellow = "\033[33m"
-	ansiReset  = "\033[0m"
-)
-
-const (
-	termChartWidth  = 70
-	termChartHeight = 12
-)
-
-// RenderGPUTerminalChart returns ANSI line charts (asciigraph-style) per GPU.
-// Used in SAT stress-test logs.
-func RenderGPUTerminalChart(rows []GPUMetricRow) string {
-	seen := make(map[int]bool)
-	var order []int
-	gpuMap := make(map[int][]GPUMetricRow)
-	for _, r := range rows {
-		if !seen[r.GPUIndex] {
-			seen[r.GPUIndex] = true
-			order = append(order, r.GPUIndex)
-		}
-		gpuMap[r.GPUIndex] = append(gpuMap[r.GPUIndex], r)
-	}
-
-	type seriesDef struct {
-		caption string
-		color   string
-		fn      func(GPUMetricRow) float64
-	}
-	defs := []seriesDef{
-		{"Temperature (°C)", ansiRed, func(r GPUMetricRow) float64 { return r.TempC }},
-		{"GPU Usage (%)", ansiBlue, func(r GPUMetricRow) float64 { return r.UsagePct }},
-		{"Power (W)", ansiGreen, func(r GPUMetricRow) float64 { return r.PowerW }},
-		{"Clock (MHz)", ansiYellow, func(r GPUMetricRow) float64 { return r.ClockMHz }},
-	}
-
-	var b strings.Builder
-	for _, gpuIdx := range order {
-		gr := gpuMap[gpuIdx]
-		if len(gr) == 0 {
-			continue
-		}
-		tMax := gr[len(gr)-1].ElapsedSec - gr[0].ElapsedSec
-		fmt.Fprintf(&b, "GPU %d — Stress Test Metrics  (%.0f seconds)\n\n", gpuIdx, tMax)
-		for _, d := range defs {
-			b.WriteString(renderLineChart(extractGPUField(gr, d.fn), d.color, d.caption,
-				termChartHeight, termChartWidth))
-			b.WriteRune('\n')
-		}
-	}
-
-	return strings.TrimRight(b.String(), "\n")
-}
-
-// renderLineChart draws a single time-series line chart using box-drawing characters.
-// Produces output in the style of asciigraph: ╭─╮ │ ╰─╯ with a Y axis and caption.
-func renderLineChart(vals []float64, color, caption string, height, width int) string {
-	if len(vals) == 0 {
-		return caption + "\n"
-	}
-
-	mn, mx := gpuMinMax(vals)
-	if mn == mx {
-		mx = mn + 1
-	}
-
-	// Use the smaller of width or len(vals) to avoid stretching sparse data.
-	w := width
-	if len(vals) < w {
-		w = len(vals)
-	}
-	data := gpuDownsample(vals, w)
-
-	// row[i] = display row index: 0 = top = max value, height = bottom = min value.
-	row := make([]int, w)
-	for i, v := range data {
-		r := int(math.Round((mx - v) / (mx - mn) * float64(height)))
-		if r < 0 {
-			r = 0
-		}
-		if r > height {
-			r = height
-		}
-		row[i] = r
-	}
-
-	// Fill the character grid.
-	grid := make([][]rune, height+1)
-	for i := range grid {
-		grid[i] = make([]rune, w)
-		for j := range grid[i] {
-			grid[i][j] = ' '
-		}
-	}
-	for x := 0; x < w; x++ {
-		r := row[x]
-		if x == 0 {
-			grid[r][0] = '─'
-			continue
-		}
-		p := row[x-1]
-		switch {
-		case r == p:
-			grid[r][x] = '─'
-		case r < p: // value went up (row index decreased toward top)
-			grid[r][x] = '╭'
-			grid[p][x] = '╯'
-			for y := r + 1; y < p; y++ {
-				grid[y][x] = '│'
-			}
-		default: // r > p, value went down
-			grid[p][x] = '╮'
-			grid[r][x] = '╰'
-			for y := p + 1; y < r; y++ {
-				grid[y][x] = '│'
-			}
-		}
-	}
-
-	// Y axis tick labels.
-	ticks := gpuNiceTicks(mn, mx, height/2)
-	tickAtRow := make(map[int]string)
-	labelWidth := 4
-	for _, t := range ticks {
-		r := int(math.Round((mx - t) / (mx - mn) * float64(height)))
-		if r < 0 || r > height {
-			continue
-		}
-		s := gpuFormatTick(t)
-		tickAtRow[r] = s
-		if len(s) > labelWidth {
-			labelWidth = len(s)
-		}
-	}
-
-	var b strings.Builder
-	for r := 0; r <= height; r++ {
-		label := tickAtRow[r]
-		fmt.Fprintf(&b, "%*s", labelWidth, label)
-		switch {
-		case label != "":
-			b.WriteRune('┤')
-		case r == height:
-			b.WriteRune('┼')
-		default:
-			b.WriteRune('│')
-		}
-		b.WriteString(color)
-		b.WriteString(string(grid[r]))
-		b.WriteString(ansiReset)
-		b.WriteRune('\n')
-	}
-
-	// Bottom axis.
-	b.WriteString(strings.Repeat(" ", labelWidth))
-	b.WriteRune('└')
-	b.WriteString(strings.Repeat("─", w))
-	b.WriteRune('\n')
-
-	// Caption centered under the chart.
-	if caption != "" {
-		total := labelWidth + 1 + w
-		if pad := (total - len(caption)) / 2; pad > 0 {
-			b.WriteString(strings.Repeat(" ", pad))
-		}
-		b.WriteString(caption)
-		b.WriteRune('\n')
-	}
-
-	return b.String()
-}
-
-func extractGPUField(rows []GPUMetricRow, fn func(GPUMetricRow) float64) []float64 {
-	v := make([]float64, len(rows))
-	for i, r := range rows {
-		v[i] = fn(r)
-	}
-	return v
-}
-
-// gpuDownsample averages vals into w buckets (or nearest-neighbor upsamples if len(vals) < w).
-func gpuDownsample(vals []float64, w int) []float64 {
-	n := len(vals)
-	if n == 0 {
-		return make([]float64, w)
-	}
-	result := make([]float64, w)
-	if n >= w {
-		counts := make([]int, w)
-		for i, v := range vals {
-			bucket := i * w / n
-			if bucket >= w {
-				bucket = w - 1
-			}
-			result[bucket] += v
-			counts[bucket]++
-		}
-		for i := range result {
-			if counts[i] > 0 {
-				result[i] /= float64(counts[i])
-			}
-		}
-	} else {
-		// Nearest-neighbour upsample.
-		for i := range result {
-			src := i * (n - 1) / (w - 1)
-			if src >= n {
-				src = n - 1
-			}
-			result[i] = vals[src]
-		}
-	}
-	return result
-}
-
 func gpuMinMax(vals []float64) (float64, float64) {
 	if len(vals) == 0 {
 		return 0, 1
@@ -644,3 +498,57 @@ func gpuFormatTick(v float64) string {
 	}
 	return strconv.FormatFloat(v, 'f', 1, 64)
 }
+
+var gpuMetricStagePalette = []string{
+	"#d95c5c",
+	"#2185d0",
+	"#21ba45",
+	"#f2c037",
+	"#6435c9",
+	"#00b5ad",
+	"#a5673f",
+}
+
+func buildGPUMetricStageSpans(rows []GPUMetricRow) []gpuMetricStageSpan {
+	var spans []gpuMetricStageSpan
+	for _, row := range rows {
+		name := strings.TrimSpace(row.Stage)
+		if name == "" {
+			name = "run"
+		}
+		start := row.StageStartSec
+		end := row.StageEndSec
+		if end <= start {
+			start = row.ElapsedSec
+			end = row.ElapsedSec
+		}
+		if len(spans) == 0 || spans[len(spans)-1].Name != name {
+			spans = append(spans, gpuMetricStageSpan{Name: name, Start: start, End: end})
+			continue
+		}
+		if start < spans[len(spans)-1].Start {
+			spans[len(spans)-1].Start = start
+		}
+		if end > spans[len(spans)-1].End {
+			spans[len(spans)-1].End = end
+		}
+	}
+	for i := range spans {
+		if spans[i].End <= spans[i].Start {
+			spans[i].End = spans[i].Start + 1
+		}
+	}
+	return spans
+}
+
+var gpuHTMLReplacer = strings.NewReplacer(
+	"&", "&amp;",
+	"<", "&lt;",
+	">", "&gt;",
+	`"`, "&quot;",
+	"'", "&#39;",
+)
+
+func gpuHTMLEscape(s string) string {
+	return gpuHTMLReplacer.Replace(s)
+}

audit/internal/platform/gpu_metrics_test.go

@@ -0,0 +1,65 @@
+package platform
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestWriteGPUMetricsCSVIncludesStageColumn(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	path := filepath.Join(dir, "gpu-metrics.csv")
+	rows := []GPUMetricRow{
+		{Stage: "warmup", ElapsedSec: 1, GPUIndex: 0, TempC: 71, UsagePct: 99, MemUsagePct: 80, PowerW: 420, ClockMHz: 1800, MemClockMHz: 1200},
+	}
+	if err := WriteGPUMetricsCSV(path, rows); err != nil {
+		t.Fatalf("WriteGPUMetricsCSV: %v", err)
+	}
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("ReadFile: %v", err)
+	}
+	text := string(raw)
+	for _, needle := range []string{
+		"stage,elapsed_sec,gpu_index",
+		`"warmup",1.0,0,71.0,99.0,80.0,420.0,1800,1200`,
+	} {
+		if !strings.Contains(text, needle) {
+			t.Fatalf("csv missing %q\n%s", needle, text)
+		}
+	}
+}
+
+func TestWriteGPUMetricsHTMLShowsStageLegendAndLabels(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	path := filepath.Join(dir, "gpu-metrics.html")
+	rows := []GPUMetricRow{
+		{Stage: "baseline", ElapsedSec: 1, GPUIndex: 0, TempC: 50, UsagePct: 10, MemUsagePct: 5, PowerW: 100, ClockMHz: 500, MemClockMHz: 400},
+		{Stage: "baseline", ElapsedSec: 2, GPUIndex: 0, TempC: 51, UsagePct: 11, MemUsagePct: 5, PowerW: 101, ClockMHz: 510, MemClockMHz: 400},
+		{Stage: "steady-fp16", ElapsedSec: 3, GPUIndex: 0, TempC: 70, UsagePct: 98, MemUsagePct: 75, PowerW: 390, ClockMHz: 1700, MemClockMHz: 1100},
+		{Stage: "steady-fp16", ElapsedSec: 4, GPUIndex: 0, TempC: 71, UsagePct: 99, MemUsagePct: 76, PowerW: 395, ClockMHz: 1710, MemClockMHz: 1110},
+	}
+	if err := WriteGPUMetricsHTML(path, rows); err != nil {
+		t.Fatalf("WriteGPUMetricsHTML: %v", err)
+	}
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("ReadFile: %v", err)
+	}
+	text := string(raw)
+	for _, needle := range []string{
+		"stage-legend",
+		"baseline",
+		"steady-fp16",
+		"GPU Stress Test Metrics",
+	} {
+		if !strings.Contains(text, needle) {
+			t.Fatalf("html missing %q\n%s", needle, text)
+		}
+	}
+}

audit/internal/platform/hpl.go

@@ -1,142 +0,0 @@
-package platform
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"strconv"
-	"strings"
-	"time"
-)
-
-// HPLOptions configures the HPL (LINPACK) benchmark run.
-type HPLOptions struct {
-	MemFraction float64 // fraction of RAM to use (default 0.80)
-	NB          int     // block size (default 256)
-}
-
-// HPLResult holds the parsed result of an HPL run.
-type HPLResult struct {
-	N          int     // matrix dimension
-	NB         int     // block size
-	P          int     // process grid rows
-	Q          int     // process grid cols
-	TimeSec    float64 // wall time in seconds
-	GFlops     float64 // achieved performance
-	Residual   float64 // backward error residual (from HPL verification line)
-	Status     string  // "PASSED" or "FAILED"
-	RawOutput  string  // full xhpl output
-}
-
-func applyHPLDefaults(opts *HPLOptions) {
-	if opts.MemFraction <= 0 || opts.MemFraction > 1 {
-		opts.MemFraction = 0.80
-	}
-	if opts.NB <= 0 {
-		opts.NB = 256
-	}
-}
-
-// RunHPL runs bee-hpl and returns parsed results plus a tar.gz artifact path.
-func (s *System) RunHPL(ctx context.Context, baseDir string, opts HPLOptions, logFunc func(string)) (string, *HPLResult, error) {
-	applyHPLDefaults(&opts)
-
-	if baseDir == "" {
-		baseDir = "/var/log/bee-sat"
-	}
-	ts := time.Now().UTC().Format("20060102-150405")
-	runDir := filepath.Join(baseDir, "hpl-"+ts)
-	if err := os.MkdirAll(runDir, 0755); err != nil {
-		return "", nil, fmt.Errorf("mkdir %s: %w", runDir, err)
-	}
-
-	logPath := filepath.Join(runDir, "hpl.log")
-
-	cmd := []string{
-		"bee-hpl",
-		"--mem-fraction", strconv.FormatFloat(opts.MemFraction, 'f', 2, 64),
-		"--nb", strconv.Itoa(opts.NB),
-	}
-
-	if logFunc != nil {
-		logFunc(fmt.Sprintf("HPL: N will be auto-sized to %.0f%% of RAM, NB=%d", opts.MemFraction*100, opts.NB))
-	}
-
-	out, err := runSATCommandCtx(ctx, "", "hpl", cmd, nil, logFunc)
-	_ = os.WriteFile(logPath, out, 0644)
-
-	result := parseHPLOutput(string(out))
-	result.RawOutput = string(out)
-
-	if err != nil && err != context.Canceled {
-		return "", result, fmt.Errorf("bee-hpl failed: %w", err)
-	}
-	if err == nil && result.GFlops <= 0 {
-		return "", result, fmt.Errorf("HPL completed but no Gflops result found in output")
-	}
-
-	// Write summary
-	summary := fmt.Sprintf("N=%d NB=%d time=%.2fs gflops=%.3f status=%s\n",
-		result.N, result.NB, result.TimeSec, result.GFlops, result.Status)
-	_ = os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary), 0644)
-
-	if logFunc != nil {
-		logFunc(fmt.Sprintf("HPL result: N=%d NB=%d %.2fs %.3f Gflops %s",
-			result.N, result.NB, result.TimeSec, result.GFlops, result.Status))
-	}
-
-	ts2 := time.Now().UTC().Format("20060102-150405")
-	archive := filepath.Join(baseDir, "hpl-"+ts2+".tar.gz")
-	if archErr := createTarGz(archive, runDir); archErr != nil {
-		return runDir, result, err
-	}
-	return archive, result, err
-}
-
-// parseHPLOutput extracts N, NB, time, and Gflops from standard HPL output.
-//
-// HPL prints a result line of the form:
-//
-//	WR00L2L2       45312   256     1     1        1234.56             5.678e+01
-//	T/V               N    NB     P     Q           Time                 Gflops
-func parseHPLOutput(output string) *HPLResult {
-	result := &HPLResult{Status: "FAILED"}
-	for _, line := range strings.Split(output, "\n") {
-		line = strings.TrimSpace(line)
-		// Result line starts with WR
-		if strings.HasPrefix(line, "WR") {
-			fields := strings.Fields(line)
-			// WR00L2L2  N  NB  P  Q  Time  Gflops
-			if len(fields) >= 7 {
-				result.N, _ = strconv.Atoi(fields[1])
-				result.NB, _ = strconv.Atoi(fields[2])
-				result.P, _ = strconv.Atoi(fields[3])
-				result.Q, _ = strconv.Atoi(fields[4])
-				result.TimeSec, _ = strconv.ParseFloat(fields[5], 64)
-				result.GFlops, _ = strconv.ParseFloat(fields[6], 64)
-			}
-		}
-		// Verification line: "||Ax-b||_oo/(eps*(||A||_oo*||x||_oo+||b||_oo)*N)= ... PASSED"
-		if strings.Contains(line, "PASSED") {
-			result.Status = "PASSED"
-			fields := strings.Fields(line)
-			for i, f := range fields {
-				if f == "PASSED" && i > 0 {
-					result.Residual, _ = strconv.ParseFloat(fields[i-1], 64)
-				}
-			}
-		}
-	}
-	return result
-}
-
-// hplAvailable returns true if bee-hpl and xhpl are present and executable.
-func hplAvailable() bool {
-	if _, err := exec.LookPath("bee-hpl"); err != nil {
-		return false
-	}
-	_, err := os.Stat("/usr/local/lib/bee/xhpl")
-	return err == nil
-}

audit/internal/platform/install_to_ram.go

@@ -11,12 +11,10 @@ import (
 	"strings"
 )
 
+const installToRAMDir = "/dev/shm/bee-live"
+
 func (s *System) IsLiveMediaInRAM() bool {
-	fsType := mountFSType("/run/live/medium")
-	if fsType == "" {
-		return toramActive()
-	}
-	return strings.EqualFold(fsType, "tmpfs")
+	return s.LiveMediaRAMState().InRAM
 }
 
 func (s *System) LiveBootSource() LiveBootSource {
@@ -48,14 +46,95 @@ func (s *System) LiveBootSource() LiveBootSource {
 	return status
 }
 
-func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) error {
+func (s *System) LiveMediaRAMState() LiveMediaRAMState {
+	return evaluateLiveMediaRAMState(
+		s.LiveBootSource(),
+		toramActive(),
+		globPaths("/run/live/medium/live/*.squashfs"),
+		globPaths(filepath.Join(installToRAMDir, "*.squashfs")),
+	)
+}
+
+func evaluateLiveMediaRAMState(status LiveBootSource, toram bool, sourceSquashfs, copiedSquashfs []string) LiveMediaRAMState {
+	state := LiveMediaRAMState{
+		LiveBootSource: status,
+		ToramActive:    toram,
+		CopyPresent:    len(copiedSquashfs) > 0,
+	}
+	if status.InRAM {
+		state.State = "in_ram"
+		state.Status = "ok"
+		state.CopyComplete = true
+		state.Message = "Running from RAM — installation media can be safely disconnected."
+		return state
+	}
+
+	expected := pathBaseSet(sourceSquashfs)
+	copied := pathBaseSet(copiedSquashfs)
+	state.CopyComplete = len(expected) > 0 && setContainsAll(copied, expected)
+
+	switch {
+	case state.CopyComplete:
+		state.State = "partial"
+		state.Status = "partial"
+		state.CanStartCopy = true
+		state.Message = "Live media files were copied to RAM, but the system is still mounted from the original boot source."
+	case state.CopyPresent:
+		state.State = "partial"
+		state.Status = "partial"
+		state.CanStartCopy = true
+		state.Message = "Partial RAM copy detected. A previous Copy to RAM run was interrupted or cancelled."
+	case toram:
+		state.State = "toram_failed"
+		state.Status = "failed"
+		state.CanStartCopy = true
+		state.Message = "toram boot parameter is set but the live medium is not mounted from RAM."
+	default:
+		state.State = "not_in_ram"
+		state.Status = "warning"
+		state.CanStartCopy = true
+		state.Message = "ISO not copied to RAM. Use Copy to RAM to free the boot drive and improve performance."
+	}
+	return state
+}
+
+func globPaths(pattern string) []string {
+	matches, _ := filepath.Glob(pattern)
+	return matches
+}
+
+func pathBaseSet(paths []string) map[string]struct{} {
+	out := make(map[string]struct{}, len(paths))
+	for _, path := range paths {
+		base := strings.TrimSpace(filepath.Base(path))
+		if base != "" {
+			out[base] = struct{}{}
+		}
+	}
+	return out
+}
+
+func setContainsAll(have, want map[string]struct{}) bool {
+	if len(want) == 0 {
+		return false
+	}
+	for name := range want {
+		if _, ok := have[name]; !ok {
+			return false
+		}
+	}
+	return true
+}
+
+func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) (retErr error) {
 	log := func(msg string) {
 		if logFunc != nil {
 			logFunc(msg)
 		}
 	}
 
-	if s.IsLiveMediaInRAM() {
+	state := s.LiveMediaRAMState()
+	if state.InRAM {
 		log("Already running from RAM — installation media can be safely disconnected.")
 		return nil
 	}
@@ -80,10 +159,21 @@ func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) erro
 			humanBytes(needed+headroom), humanBytes(free))
 	}
 
-	dstDir := "/dev/shm/bee-live"
+	dstDir := installToRAMDir
+	if state.CopyPresent {
+		log("Removing stale partial RAM copy before retry...")
+	}
+	_ = os.RemoveAll(dstDir)
 	if err := os.MkdirAll(dstDir, 0755); err != nil {
 		return fmt.Errorf("create tmpfs dir: %v", err)
 	}
+	defer func() {
+		if retErr == nil {
+			return
+		}
+		_ = os.RemoveAll(dstDir)
+		log("Removed incomplete RAM copy.")
+	}()
 
 	for _, sf := range squashfsFiles {
 		if err := ctx.Err(); err != nil {

audit/internal/platform/install_to_ram_test.go

@@ -58,3 +58,46 @@ func TestDescribeLiveBootSource(t *testing.T) {
 		t.Fatalf("got %q want /run/live/medium", got)
 	}
 }
+
+func TestEvaluateLiveMediaRAMState(t *testing.T) {
+	t.Parallel()
+
+	t.Run("in_ram", func(t *testing.T) {
+		state := evaluateLiveMediaRAMState(
+			LiveBootSource{InRAM: true, Kind: "ram", Source: "tmpfs"},
+			false,
+			nil,
+			nil,
+		)
+		if state.State != "in_ram" || state.Status != "ok" || state.CanStartCopy {
+			t.Fatalf("state=%+v", state)
+		}
+	})
+
+	t.Run("partial_copy_after_cancel", func(t *testing.T) {
+		state := evaluateLiveMediaRAMState(
+			LiveBootSource{InRAM: false, Kind: "usb", Device: "/dev/sdb1"},
+			false,
+			[]string{"/run/live/medium/live/filesystem.squashfs", "/run/live/medium/live/firmware.squashfs"},
+			[]string{"/dev/shm/bee-live/filesystem.squashfs"},
+		)
+		if state.State != "partial" || state.Status != "partial" || !state.CanStartCopy {
+			t.Fatalf("state=%+v", state)
+		}
+		if state.CopyComplete {
+			t.Fatalf("CopyComplete=%v want false", state.CopyComplete)
+		}
+	})
+
+	t.Run("toram_failed", func(t *testing.T) {
+		state := evaluateLiveMediaRAMState(
+			LiveBootSource{InRAM: false, Kind: "usb", Device: "/dev/sdb1"},
+			true,
+			nil,
+			nil,
+		)
+		if state.State != "toram_failed" || state.Status != "failed" || !state.CanStartCopy {
+			t.Fatalf("state=%+v", state)
+		}
+	})
+}

audit/internal/platform/nvidia_stress.go

@@ -49,6 +49,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"--seconds", strconv.Itoa(opts.DurationSec),
 			"--size-mb", strconv.Itoa(opts.SizeMB),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}
@@ -63,6 +66,9 @@ func buildNvidiaStressJob(opts NvidiaStressOptions) (satJob, error) {
 			"bee-john-gpu-stress",
 			"--seconds", strconv.Itoa(opts.DurationSec),
 		}
+		if opts.StaggerSeconds > 0 && len(selected) > 1 {
+			cmd = append(cmd, "--stagger-seconds", strconv.Itoa(opts.StaggerSeconds))
+		}
 		if len(selected) > 0 {
 			cmd = append(cmd, "--devices", joinIndexList(selected))
 		}

audit/internal/platform/platform_stress.go

@@ -161,13 +161,7 @@ func (s *System) RunPlatformStress(
 	}
 	_ = os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary), 0644)
 
-	// Pack tar.gz
-	archivePath := filepath.Join(baseDir, "platform-stress-"+stamp+".tar.gz")
-	if err := packPlatformDir(runDir, archivePath); err != nil {
-		return "", fmt.Errorf("pack archive: %w", err)
-	}
-	_ = os.RemoveAll(runDir)
-	return archivePath, nil
+	return runDir, nil
 }
 
 // collectPhase samples live metrics every second until ctx is done.

audit/internal/platform/runtime.go

@@ -1,6 +1,7 @@
 package platform
 
 import (
+	"bufio"
 	"os"
 	"os/exec"
 	"strings"
@@ -27,6 +28,8 @@ var runtimeTrackedServices = []string{
 	"bee-audit",
 	"bee-web",
 	"bee-sshsetup",
+	"nvidia-dcgm",
+	"nvidia-fabricmanager",
 }
 
 func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, error) {
@@ -114,6 +117,8 @@ func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, e
 	}
 
 	s.collectGPURuntimeHealth(vendor, &health)
+	s.collectToRAMHealth(&health)
+	s.collectUSBExportHealth(&health)
 
 	if health.Status != "FAILED" && len(health.Issues) > 0 {
 		health.Status = "PARTIAL"
@@ -168,6 +173,99 @@ func resolvedToolStatus(display string, candidates ...string) ToolStatus {
 	return ToolStatus{Name: display}
 }
 
+// collectToRAMHealth evaluates whether the live system is fully running from RAM.
+// Status values: "ok" = fully in RAM, "warning" = not copied, "partial" = stale or
+// incomplete RAM copy exists but runtime still depends on the boot medium,
+// "failed" = toram was requested but medium is not in RAM.
+func (s *System) collectToRAMHealth(health *schema.RuntimeHealth) {
+	state := s.LiveMediaRAMState()
+	health.ToRAMStatus = state.Status
+	switch state.Status {
+	case "ok":
+		return
+	case "failed":
+		health.Issues = append(health.Issues, schema.RuntimeIssue{
+			Code:        "toram_copy_failed",
+			Severity:    "warning",
+			Description: state.Message,
+		})
+	case "partial":
+		health.Issues = append(health.Issues, schema.RuntimeIssue{
+			Code:        "toram_copy_partial",
+			Severity:    "warning",
+			Description: state.Message,
+		})
+	}
+}
+
+// collectUSBExportHealth scans /proc/mounts for a writable USB-backed filesystem
+// suitable for log export. Sets USBExportPath to the first match found.
+func (s *System) collectUSBExportHealth(health *schema.RuntimeHealth) {
+	health.USBExportPath = findUSBExportMount()
+}
+
+// findUSBExportMount returns the mount point of the first writable USB filesystem
+// found in /proc/mounts (vfat, exfat, ext2/3/4, ntfs) whose backing block device
+// has USB transport. Returns "" if none found.
+func findUSBExportMount() string {
+	f, err := os.Open("/proc/mounts")
+	if err != nil {
+		return ""
+	}
+	defer f.Close()
+
+	// fs types that are expected on USB export drives
+	exportFSTypes := map[string]bool{
+		"vfat":    true,
+		"exfat":   true,
+		"ext2":    true,
+		"ext3":    true,
+		"ext4":    true,
+		"ntfs":    true,
+		"ntfs3":   true,
+		"fuseblk": true,
+	}
+
+	scanner := bufio.NewScanner(f)
+	for scanner.Scan() {
+		// fields: device mountpoint fstype options dump pass
+		fields := strings.Fields(scanner.Text())
+		if len(fields) < 4 {
+			continue
+		}
+		device, mountPoint, fsType, options := fields[0], fields[1], fields[2], fields[3]
+		if !exportFSTypes[strings.ToLower(fsType)] {
+			continue
+		}
+		// Skip read-only mounts
+		opts := strings.Split(options, ",")
+		readOnly := false
+		for _, o := range opts {
+			if strings.TrimSpace(o) == "ro" {
+				readOnly = true
+				break
+			}
+		}
+		if readOnly {
+			continue
+		}
+		// Check USB transport via lsblk on the device (or its parent disk for partitions).
+		if !strings.HasPrefix(device, "/dev/") {
+			continue
+		}
+		checkDev := device
+		// lsblk only reports TRAN for the whole disk, not for partitions (e.g. /dev/sdc1).
+		// Strip trailing partition digits to get the parent disk name.
+		if trimmed := strings.TrimRight(device, "0123456789"); trimmed != device && len(trimmed) > len("/dev/") {
+			checkDev = trimmed
+		}
+		if blockDeviceTransport(checkDev) == "usb" {
+			return mountPoint
+		}
+	}
+	return ""
+}
+
 func (s *System) collectGPURuntimeHealth(vendor string, health *schema.RuntimeHealth) {
 	lsmodText := commandText("lsmod")
 

audit/internal/platform/sat.go

@@ -108,15 +108,15 @@ type nvidiaGPUHealth struct {
 }
 
 type nvidiaGPUStatusFile struct {
-	Index       int
-	Name        string
-	RunStatus   string
-	Reason      string
-	Health      string
-	HealthRaw   string
-	Observed    bool
-	Selected    bool
-	FailingJob  string
+	Index      int
+	Name       string
+	RunStatus  string
+	Reason     string
+	Health     string
+	HealthRaw  string
+	Observed   bool
+	Selected   bool
+	FailingJob string
 }
 
 // AMDGPUInfo holds basic info about an AMD GPU from rocm-smi.
@@ -384,14 +384,28 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 	), logFunc)
 }
 
-func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
+func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, staggerSec int, logFunc func(string)) (string, error) {
 	selected, err := resolveDCGMGPUIndices(gpuIndices)
 	if err != nil {
 		return "", err
 	}
-	profCmd, err := resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004", "-d", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)))
-	if err != nil {
-		return "", err
+	var (
+		profCmd []string
+		profEnv []string
+	)
+	if staggerSec > 0 && len(selected) > 1 {
+		profCmd = []string{
+			"bee-dcgmproftester-staggered",
+			"--seconds", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)),
+			"--stagger-seconds", strconv.Itoa(staggerSec),
+			"--devices", joinIndexList(selected),
+		}
+	} else {
+		profCmd, err = resolveDCGMProfTesterCommand("--no-dcgm-validation", "-t", "1004", "-d", strconv.Itoa(normalizeNvidiaBurnDuration(durationSec)))
+		if err != nil {
+			return "", err
+		}
+		profEnv = nvidiaVisibleDevicesEnv(selected)
 	}
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
@@ -399,7 +413,7 @@ func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir strin
 		satJob{
 			name:       "03-dcgmproftester.log",
 			cmd:        profCmd,
-			env:        nvidiaVisibleDevicesEnv(selected),
+			env:        profEnv,
 			collectGPU: true,
 			gpuIndices: selected,
 		},
@@ -412,6 +426,13 @@ func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string,
 	if err != nil {
 		return "", err
 	}
+	// Kill any lingering nvvs/dcgmi processes from a previous interrupted run
+	// before starting — otherwise dcgmi diag fails with DCGM_ST_IN_USE (-34).
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{
@@ -429,6 +450,13 @@ func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, dur
 	if err != nil {
 		return "", err
 	}
+	// Kill any lingering nvvs/dcgmi processes from a previous interrupted run
+	// before starting — otherwise dcgmi diag fails with DCGM_ST_IN_USE (-34).
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{
@@ -446,6 +474,13 @@ func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpu
 	if err != nil {
 		return "", err
 	}
+	// Kill any lingering nvvs/dcgmi processes from a previous interrupted run
+	// before starting — otherwise dcgmi diag fails with DCGM_ST_IN_USE (-34).
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{
@@ -538,9 +573,19 @@ func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, si
 	if passes <= 0 {
 		passes = 1
 	}
+	// Keep Validate Memory bounded to a quick diagnostic window. The timeout is
+	// intentionally conservative enough for healthy systems while avoiding the
+	// prior 30-80 minute hangs caused by memtester spinning on a bad subtest.
+	timeoutSec := sizeMB*passes*20/100 + 60
+	if timeoutSec < 180 {
+		timeoutSec = 180
+	}
+	if timeoutSec > 900 {
+		timeoutSec = 900
+	}
 	return runAcceptancePackCtx(ctx, baseDir, "memory", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
-		{name: "02-memtester.log", cmd: []string{"memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
+		{name: "02-memtester.log", cmd: []string{"timeout", fmt.Sprintf("%d", timeoutSec), "memtester", fmt.Sprintf("%dM", sizeMB), fmt.Sprintf("%d", passes)}},
 		{name: "03-free-after.log", cmd: []string{"free", "-h"}},
 	}, logFunc)
 }
@@ -648,11 +693,7 @@ func (s *System) RunStorageAcceptancePack(ctx context.Context, baseDir string, e
 	if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte(summary.String()), 0644); err != nil {
 		return "", err
 	}
-	archive := filepath.Join(baseDir, "storage-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 type satJob struct {
@@ -838,11 +879,7 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		}
 	}
 
-	archive := filepath.Join(baseDir, prefix+"-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 func updateNvidiaGPUStatus(perGPU map[int]*nvidiaGPUStatusFile, idx int, status, jobName, detail string) {
@@ -905,7 +942,7 @@ func writeNvidiaGPUStatusFiles(runDir, overall string, perGPU map[int]*nvidiaGPU
 			entry.Health = "UNKNOWN"
 		}
 		if entry.Name == "" {
-			entry.Name = "unknown"
+			entry.Name = "Unknown GPU"
 		}
 		var body strings.Builder
 		fmt.Fprintf(&body, "gpu_index=%d\n", entry.Index)
@@ -1376,8 +1413,6 @@ func runSATCommandWithMetrics(ctx context.Context, verboseLog, name string, cmd
 	if len(metricRows) > 0 {
 		_ = WriteGPUMetricsCSV(filepath.Join(runDir, "gpu-metrics.csv"), metricRows)
 		_ = WriteGPUMetricsHTML(filepath.Join(runDir, "gpu-metrics.html"), metricRows)
-		chart := RenderGPUTerminalChart(metricRows)
-		_ = os.WriteFile(filepath.Join(runDir, "gpu-metrics-term.txt"), []byte(chart), 0644)
 	}
 
 	return out, err

audit/internal/platform/sat_fan_stress.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"math"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -56,13 +57,37 @@ type cachedPowerReading struct {
 	UpdatedAt time.Time
 }
 
+type fanObservationState struct {
+	MaxRPM map[string]float64 `json:"max_rpm"`
+}
+
+type fanPeakCandidate struct {
+	FirstSeen time.Time
+	RPM       float64
+}
+
 var (
 	systemPowerCacheMu sync.Mutex
 	systemPowerCache   cachedPowerReading
+	fanObservationMu   sync.Mutex
+	fanObservation     fanObservationState
+	fanObservationInit bool
+	fanPeakCandidates  = make(map[string]fanPeakCandidate)
 )
 
 const systemPowerHoldTTL = 15 * time.Second
 
+var fanObservationStatePath = "/var/log/bee-sat/fan-observation.json"
+
+const fanObservationMinPeakHold = time.Second
+
+func normalizeObservedFanMaxRPM(rpm float64) float64 {
+	if rpm <= 0 {
+		return 0
+	}
+	return math.Ceil(rpm/1000.0) * 1000.0
+}
+
 // RunFanStressTest runs a two-phase GPU stress test while monitoring fan speeds,
 // temperatures, and power draw every second. Exports metrics.csv and fan-sensors.csv.
 // Designed to reproduce case-04 fan-speed lag and detect GPU thermal throttling.
@@ -223,11 +248,7 @@ func (s *System) RunFanStressTest(ctx context.Context, baseDir string, opts FanS
 		return "", err
 	}
 
-	archive := filepath.Join(baseDir, "fan-stress-"+ts+".tar.gz")
-	if err := createTarGz(archive, runDir); err != nil {
-		return "", err
-	}
-	return archive, nil
+	return runDir, nil
 }
 
 func applyFanStressDefaults(opts *FanStressOptions) {
@@ -314,11 +335,13 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	out, err := exec.Command("ipmitool", "sdr", "type", "Fan").Output()
 	if err == nil {
 		if fans := parseFanSpeeds(string(out)); len(fans) > 0 {
+			updateFanObservation(fans, time.Now())
 			return fans, nil
 		}
 	}
 	fans, sensorsErr := sampleFanSpeedsViaSensorsJSON()
 	if len(fans) > 0 {
+		updateFanObservation(fans, time.Now())
 		return fans, nil
 	}
 	if err != nil {
@@ -327,6 +350,119 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	return nil, sensorsErr
 }
 
+func loadFanObservationLocked() {
+	if fanObservationInit {
+		return
+	}
+	fanObservationInit = true
+	fanObservation.MaxRPM = make(map[string]float64)
+	raw, err := os.ReadFile(fanObservationStatePath)
+	if err != nil || len(raw) == 0 {
+		return
+	}
+	var persisted fanObservationState
+	if json.Unmarshal(raw, &persisted) != nil {
+		return
+	}
+	for name, rpm := range persisted.MaxRPM {
+		name = strings.TrimSpace(name)
+		if name == "" || rpm <= 0 {
+			continue
+		}
+		fanObservation.MaxRPM[name] = rpm
+	}
+}
+
+func saveFanObservationLocked() {
+	if len(fanObservation.MaxRPM) == 0 {
+		return
+	}
+	dir := filepath.Dir(fanObservationStatePath)
+	if dir == "" || dir == "." {
+		dir = "/var/log/bee-sat"
+	}
+	if err := os.MkdirAll(dir, 0755); err != nil {
+		return
+	}
+	raw, err := json.MarshalIndent(fanObservation, "", "  ")
+	if err != nil {
+		return
+	}
+	_ = os.WriteFile(fanObservationStatePath, raw, 0644)
+}
+
+func updateFanObservation(fans []FanReading, now time.Time) {
+	if len(fans) == 0 {
+		return
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	changed := false
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		currentMax := fanObservation.MaxRPM[name]
+		if fan.RPM <= currentMax {
+			delete(fanPeakCandidates, name)
+			continue
+		}
+		if cand, ok := fanPeakCandidates[name]; ok {
+			if now.Sub(cand.FirstSeen) >= fanObservationMinPeakHold {
+				newMax := math.Max(cand.RPM, fan.RPM)
+				if newMax > currentMax {
+					fanObservation.MaxRPM[name] = normalizeObservedFanMaxRPM(newMax)
+					changed = true
+				}
+				delete(fanPeakCandidates, name)
+				continue
+			}
+			if fan.RPM > cand.RPM {
+				fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: cand.FirstSeen, RPM: fan.RPM}
+			}
+			continue
+		}
+		fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: now, RPM: fan.RPM}
+	}
+	if changed {
+		saveFanObservationLocked()
+	}
+}
+
+func estimateFanDutyCyclePctFromObservation(fans []FanReading) (float64, bool) {
+	if len(fans) == 0 {
+		return 0, false
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	var samples []float64
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		maxRPM := fanObservation.MaxRPM[name]
+		if maxRPM <= 0 {
+			continue
+		}
+		pct := fan.RPM / maxRPM * 100.0
+		if pct > 100 {
+			pct = 100
+		}
+		if pct < 0 {
+			pct = 0
+		}
+		samples = append(samples, pct)
+	}
+	if len(samples) == 0 {
+		return 0, false
+	}
+	return benchmarkMean(samples), true
+}
+
 // parseFanSpeeds parses "ipmitool sdr type Fan" output.
 // Handles two formats:
 //
@@ -430,6 +566,116 @@ func sampleFanSpeedsViaSensorsJSON() ([]FanReading, error) {
 	return fans, nil
 }
 
+// sampleFanDutyCyclePct reads fan PWM/duty-cycle controls from lm-sensors.
+// Returns the average duty cycle across all exposed PWM controls.
+func sampleFanDutyCyclePct() (float64, bool, bool) {
+	out, err := exec.Command("sensors", "-j").Output()
+	if err != nil || len(out) == 0 {
+		fans, fanErr := sampleFanSpeeds()
+		if fanErr != nil {
+			return 0, false, false
+		}
+		return sampleFanDutyCyclePctFromFans(fans)
+	}
+	pct, ok := parseFanDutyCyclePctSensorsJSON(out)
+	return pct, ok, false
+}
+
+func sampleFanDutyCyclePctFromFans(fans []FanReading) (float64, bool, bool) {
+	if len(fans) == 0 {
+		return 0, false, false
+	}
+	if pct, ok := estimateFanDutyCyclePctFromObservation(fans); ok {
+		return pct, true, true
+	}
+	return 0, false, false
+}
+
+func parseFanDutyCyclePctSensorsJSON(raw []byte) (float64, bool) {
+	var doc map[string]map[string]any
+	if err := json.Unmarshal(raw, &doc); err != nil {
+		return 0, false
+	}
+	var samples []float64
+	for _, features := range doc {
+		for name, feature := range features {
+			if strings.EqualFold(name, "Adapter") {
+				continue
+			}
+			featureMap, ok := feature.(map[string]any)
+			if !ok {
+				continue
+			}
+			if duty, ok := firstFanDutyValue(name, featureMap); ok {
+				samples = append(samples, duty)
+			}
+		}
+	}
+	if len(samples) == 0 {
+		return 0, false
+	}
+	return benchmarkMean(samples), true
+}
+
+func firstFanDutyValue(featureName string, feature map[string]any) (float64, bool) {
+	featureName = strings.ToLower(strings.TrimSpace(featureName))
+	if strings.Contains(featureName, "enable") || strings.Contains(featureName, "mode") || strings.Contains(featureName, "alarm") {
+		return 0, false
+	}
+	if strings.Contains(featureName, "pwm") {
+		for _, key := range []string{"input", "value", "current"} {
+			if value, ok := feature[key]; ok {
+				if duty, parsed := parseFanDutyValue(value); parsed {
+					return duty, true
+				}
+			}
+		}
+	}
+	keys := make([]string, 0, len(feature))
+	for key := range feature {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	for _, key := range keys {
+		lower := strings.ToLower(key)
+		if !strings.Contains(lower, "pwm") {
+			continue
+		}
+		if strings.Contains(lower, "enable") || strings.Contains(lower, "mode") || strings.Contains(lower, "alarm") {
+			continue
+		}
+		if duty, parsed := parseFanDutyValue(feature[key]); parsed {
+			return duty, true
+		}
+	}
+	return 0, false
+}
+
+func parseFanDutyValue(value any) (float64, bool) {
+	switch v := value.(type) {
+	case float64:
+		return normalizePWMAsDutyPct(v)
+	case string:
+		if f, err := strconv.ParseFloat(strings.TrimSpace(v), 64); err == nil {
+			return normalizePWMAsDutyPct(f)
+		}
+	}
+	return 0, false
+}
+
+func normalizePWMAsDutyPct(raw float64) (float64, bool) {
+	if raw < 0 {
+		return 0, false
+	}
+	if raw <= 100 {
+		return raw, true
+	}
+	if raw <= 255 {
+		return raw / 255.0 * 100.0, true
+	}
+	return 0, false
+}
+
 func firstFanInputValue(feature map[string]any) (float64, bool) {
 	keys := make([]string, 0, len(feature))
 	for key := range feature {

audit/internal/platform/sat_fan_stress_test.go

@@ -1,6 +1,7 @@
 package platform
 
 import (
+	"path/filepath"
 	"testing"
 	"time"
 )
@@ -29,6 +30,74 @@ func TestFirstFanInputValue(t *testing.T) {
 	}
 }
 
+func TestParseFanDutyCyclePctSensorsJSON(t *testing.T) {
+	raw := []byte(`{
+		"chip0": {
+			"fan1": {"input": 9000},
+			"pwm1": {"input": 128},
+			"pwm1_enable": {"input": 1}
+		},
+		"chip1": {
+			"pwm2": {"input": 64}
+		}
+	}`)
+
+	got, ok := parseFanDutyCyclePctSensorsJSON(raw)
+	if !ok {
+		t.Fatalf("expected duty cycle telemetry to be parsed")
+	}
+	if got < 57 || got > 58 {
+		t.Fatalf("got=%v want ~57.1", got)
+	}
+}
+
+func TestEstimateFanDutyCyclePctFromObservation(t *testing.T) {
+	t.Parallel()
+
+	oldPath := fanObservationStatePath
+	oldState := fanObservation
+	oldInit := fanObservationInit
+	oldCandidates := fanPeakCandidates
+	fanObservationStatePath = filepath.Join(t.TempDir(), "fan-observation.json")
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	t.Cleanup(func() {
+		fanObservationStatePath = oldPath
+		fanObservation = oldState
+		fanObservationInit = oldInit
+		fanPeakCandidates = oldCandidates
+	})
+
+	start := time.Unix(100, 0)
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5000}}, start)
+	if _, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2500}}); ok {
+		t.Fatalf("single-sample spike should not establish observed max")
+	}
+
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5200}}, start.Add(500*time.Millisecond))
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5100}}, start.Add(1500*time.Millisecond))
+
+	got, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected estimated duty cycle from persisted observed max")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("got=%v want ~43.3", got)
+	}
+
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	got, ok = estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected persisted observed max to be reloaded from disk")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("reloaded got=%v want ~43.3", got)
+	}
+}
+
 func TestParseDCMIPowerReading(t *testing.T) {
 	raw := `
 Instantaneous power reading:                   512 Watts

audit/internal/platform/techdump.go

@@ -20,6 +20,7 @@ var techDumpFixedCommands = []struct {
 	{Name: "dmidecode", Args: []string{"-t", "4"}, File: "dmidecode-type4.txt"},
 	{Name: "dmidecode", Args: []string{"-t", "17"}, File: "dmidecode-type17.txt"},
 	{Name: "lspci", Args: []string{"-vmm", "-D"}, File: "lspci-vmm.txt"},
+	{Name: "lspci", Args: []string{"-vvv"}, File: "lspci-vvv.txt"},
 	{Name: "lsblk", Args: []string{"-J", "-d", "-o", "NAME,TYPE,SIZE,SERIAL,MODEL,TRAN,HCTL"}, File: "lsblk.json"},
 	{Name: "sensors", Args: []string{"-j"}, File: "sensors.json"},
 	{Name: "ipmitool", Args: []string{"fru", "print"}, File: "ipmitool-fru.txt"},

audit/internal/platform/types.go

@@ -9,6 +9,17 @@ type LiveBootSource struct {
 	Device string `json:"device,omitempty"`
 }
 
+type LiveMediaRAMState struct {
+	LiveBootSource
+	State        string `json:"state"`
+	Status       string `json:"status"`
+	ToramActive  bool   `json:"toram_active,omitempty"`
+	CopyPresent  bool   `json:"copy_present,omitempty"`
+	CopyComplete bool   `json:"copy_complete,omitempty"`
+	CanStartCopy bool   `json:"can_start_copy,omitempty"`
+	Message      string `json:"message,omitempty"`
+}
+
 type InterfaceInfo struct {
 	Name  string
 	State string
@@ -70,6 +81,7 @@ type NvidiaStressOptions struct {
 	Loader            string
 	GPUIndices        []int
 	ExcludeGPUIndices []int
+	StaggerSeconds    int
 }
 
 func New() *System {

audit/internal/schema/hardware.go

@@ -15,13 +15,17 @@ type HardwareIngestRequest struct {
 }
 
 type RuntimeHealth struct {
-	Status        string                 `json:"status"`
-	CheckedAt     string                 `json:"checked_at"`
-	ExportDir     string                 `json:"export_dir,omitempty"`
-	DriverReady   bool                   `json:"driver_ready,omitempty"`
-	CUDAReady     bool                   `json:"cuda_ready,omitempty"`
-	NvidiaGSPMode string                 `json:"nvidia_gsp_mode,omitempty"` // "gsp-on", "gsp-off", "gsp-stuck"
-	NetworkStatus string                 `json:"network_status,omitempty"`
+	Status        string `json:"status"`
+	CheckedAt     string `json:"checked_at"`
+	ExportDir     string `json:"export_dir,omitempty"`
+	DriverReady   bool   `json:"driver_ready,omitempty"`
+	CUDAReady     bool   `json:"cuda_ready,omitempty"`
+	NvidiaGSPMode string `json:"nvidia_gsp_mode,omitempty"` // "gsp-on", "gsp-off", "gsp-stuck"
+	NetworkStatus string `json:"network_status,omitempty"`
+	// ToRAMStatus: "ok" (fully in RAM), "warning" (not copied), "partial" (stale/incomplete copy exists), "failed" (toram active but copy failed)
+	ToRAMStatus string `json:"toram_status,omitempty"`
+	// USBExportPath: mount point of the first writable USB drive found, empty if none.
+	USBExportPath string                 `json:"usb_export_path,omitempty"`
 	Issues        []RuntimeIssue         `json:"issues,omitempty"`
 	Tools         []RuntimeToolStatus    `json:"tools,omitempty"`
 	Services      []RuntimeServiceStatus `json:"services,omitempty"`
@@ -183,6 +187,13 @@ type HardwarePCIeDevice struct {
 	BatteryTemperatureC    *float64       `json:"battery_temperature_c,omitempty"`
 	BatteryVoltageV        *float64       `json:"battery_voltage_v,omitempty"`
 	BatteryReplaceRequired *bool          `json:"battery_replace_required,omitempty"`
+	SFPPresent             *bool          `json:"sfp_present,omitempty"`
+	SFPIdentifier          *string        `json:"sfp_identifier,omitempty"`
+	SFPConnector           *string        `json:"sfp_connector,omitempty"`
+	SFPVendor              *string        `json:"sfp_vendor,omitempty"`
+	SFPPartNumber          *string        `json:"sfp_part_number,omitempty"`
+	SFPSerialNumber        *string        `json:"sfp_serial_number,omitempty"`
+	SFPWavelengthNM        *float64       `json:"sfp_wavelength_nm,omitempty"`
 	SFPTemperatureC        *float64       `json:"sfp_temperature_c,omitempty"`
 	SFPTXPowerDBM          *float64       `json:"sfp_tx_power_dbm,omitempty"`
 	SFPRXPowerDBM          *float64       `json:"sfp_rx_power_dbm,omitempty"`

audit/internal/webui/api.go

@@ -12,6 +12,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -35,6 +36,16 @@ var apiListNvidiaGPUStatuses = func(a *app.App) ([]platform.NvidiaGPUStatus, err
 	return a.ListNvidiaGPUStatuses()
 }
 
+const (
+	taskPriorityBenchmark      = 10
+	taskPriorityBurn           = 20
+	taskPriorityValidateStress = 30
+	taskPriorityValidate       = 40
+	taskPriorityAudit          = 50
+	taskPriorityInstallToRAM   = 60
+	taskPriorityInstall        = 70
+)
+
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
@@ -99,7 +110,7 @@ func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {
 
 func shouldSplitHomogeneousNvidiaTarget(target string) bool {
 	switch strings.TrimSpace(target) {
-	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
+	case "nvidia", "nvidia-targeted-stress", "nvidia-bench-perf", "nvidia-bench-power", "nvidia-compute",
 		"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
 		"nvidia-bandwidth", "nvidia-stress":
 		return true
@@ -108,6 +119,30 @@ func shouldSplitHomogeneousNvidiaTarget(target string) bool {
 	}
 }
 
+func defaultTaskPriority(target string, params taskParams) int {
+	switch strings.TrimSpace(target) {
+	case "install":
+		return taskPriorityInstall
+	case "install-to-ram":
+		return taskPriorityInstallToRAM
+	case "audit":
+		return taskPriorityAudit
+	case "nvidia-bench-perf", "nvidia-bench-power":
+		return taskPriorityBenchmark
+	case "nvidia-stress", "amd-stress", "memory-stress", "sat-stress", "platform-stress", "nvidia-compute":
+		return taskPriorityBurn
+	case "nvidia", "nvidia-targeted-stress", "nvidia-targeted-power", "nvidia-pulse",
+		"nvidia-interconnect", "nvidia-bandwidth", "memory", "storage", "cpu",
+		"amd", "amd-mem", "amd-bandwidth":
+		if params.StressMode {
+			return taskPriorityValidateStress
+		}
+		return taskPriorityValidate
+	default:
+		return 0
+	}
+}
+
 func expandHomogeneousNvidiaSelections(gpus []platform.NvidiaGPU, include, exclude []int) ([]nvidiaTaskSelection, error) {
 	if len(gpus) == 0 {
 		return nil, fmt.Errorf("no NVIDIA GPUs detected")
@@ -209,6 +244,14 @@ func joinTaskIndices(indices []int) string {
 	return strings.Join(parts, ",")
 }
 
+func formatGPUIndexList(indices []int) string {
+	parts := make([]string, len(indices))
+	for i, idx := range indices {
+		parts[i] = strconv.Itoa(idx)
+	}
+	return strings.Join(parts, ",")
+}
+
 func formatSplitTaskName(baseName, selectionLabel string) string {
 	baseName = strings.TrimSpace(baseName)
 	selectionLabel = strings.TrimSpace(selectionLabel)
@@ -449,6 +492,7 @@ func (h *handler) handleAPIAuditRun(w http.ResponseWriter, _ *http.Request) {
 		ID:        newJobID("audit"),
 		Name:      "Audit",
 		Target:    "audit",
+		Priority:  defaultTaskPriority("audit", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 	}
@@ -487,6 +531,8 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			StressMode         bool     `json:"stress_mode"`
 			GPUIndices         []int    `json:"gpu_indices"`
 			ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
+			StaggerGPUStart    bool     `json:"stagger_gpu_start"`
+			ParallelGPUs       bool     `json:"parallel_gpus"`
 			Loader             string   `json:"loader"`
 			Profile            string   `json:"profile"`
 			DisplayName        string   `json:"display_name"`
@@ -508,12 +554,142 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			StressMode:         body.StressMode,
 			GPUIndices:         body.GPUIndices,
 			ExcludeGPUIndices:  body.ExcludeGPUIndices,
+			StaggerGPUStart:    body.StaggerGPUStart,
+			ParallelGPUs:       body.ParallelGPUs,
 			Loader:             body.Loader,
 			BurnProfile:        body.Profile,
 			DisplayName:        body.DisplayName,
 			PlatformComponents: body.PlatformComponents,
 		}
-		tasks, err := buildNvidiaTaskSet(target, 0, time.Now(), params, name, h.opts.App, "sat-"+target)
+		tasks, err := buildNvidiaTaskSet(target, defaultTaskPriority(target, params), time.Now(), params, name, h.opts.App, "sat-"+target)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
+	}
+}
+
+func (h *handler) handleAPIBenchmarkNvidiaRunKind(target string) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		if h.opts.App == nil {
+			writeError(w, http.StatusServiceUnavailable, "app not configured")
+			return
+		}
+
+		var body struct {
+			Profile           string `json:"profile"`
+			SizeMB            int    `json:"size_mb"`
+			GPUIndices        []int  `json:"gpu_indices"`
+			ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
+			RunNCCL           *bool  `json:"run_nccl"`
+			ParallelGPUs      *bool  `json:"parallel_gpus"`
+			RampUp            *bool  `json:"ramp_up"`
+			DisplayName       string `json:"display_name"`
+		}
+		if r.Body != nil {
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
+				writeError(w, http.StatusBadRequest, "invalid request body")
+				return
+			}
+		}
+
+		runNCCL := true
+		if body.RunNCCL != nil {
+			runNCCL = *body.RunNCCL
+		}
+		parallelGPUs := false
+		if body.ParallelGPUs != nil {
+			parallelGPUs = *body.ParallelGPUs
+		}
+		rampUp := false
+		if body.RampUp != nil {
+			rampUp = *body.RampUp
+		}
+		// Build a descriptive base name that includes profile and mode so the task
+		// list is self-explanatory without opening individual task detail pages.
+		profile := strings.TrimSpace(body.Profile)
+		if profile == "" {
+			profile = "standard"
+		}
+		name := taskDisplayName(target, "", "")
+		if strings.TrimSpace(body.DisplayName) != "" {
+			name = body.DisplayName
+		}
+		// Append profile tag.
+		name = fmt.Sprintf("%s · %s", name, profile)
+
+		if target == "nvidia-bench-power" && parallelGPUs {
+			writeError(w, http.StatusBadRequest, "power / thermal fit benchmark uses sequential or ramp-up modes only")
+			return
+		}
+
+		if rampUp && len(body.GPUIndices) > 1 {
+			// Ramp-up mode: RunNvidiaPowerBench internally ramps from 1 to N GPUs
+			// in Phase 2 (one additional GPU per step). A single task with all
+			// selected GPUs is sufficient — spawning N tasks with growing subsets
+			// would repeat all earlier steps redundantly.
+			gpus, err := apiListNvidiaGPUs(h.opts.App)
+			if err != nil {
+				writeError(w, http.StatusBadRequest, err.Error())
+				return
+			}
+			resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
+			if err != nil {
+				writeError(w, http.StatusBadRequest, err.Error())
+				return
+			}
+			if len(resolved) < 2 {
+				// Fall through to normal single-task path.
+				rampUp = false
+			} else {
+				now := time.Now()
+				rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
+				taskName := fmt.Sprintf("%s · ramp 1–%d · GPU %s", name, len(resolved), formatGPUIndexList(resolved))
+				t := &Task{
+					ID:        newJobID("bee-bench-nvidia"),
+					Name:      taskName,
+					Target:    target,
+					Priority:  defaultTaskPriority(target, taskParams{}),
+					Status:    TaskPending,
+					CreatedAt: now,
+					params: taskParams{
+						GPUIndices:       append([]int(nil), resolved...),
+						SizeMB:           body.SizeMB,
+						BenchmarkProfile: body.Profile,
+						RunNCCL:          runNCCL,
+						ParallelGPUs:     true,
+						RampTotal:        len(resolved),
+						RampRunID:        rampRunID,
+						DisplayName:      taskName,
+					},
+				}
+				globalQueue.enqueue(t)
+				writeTaskRunResponse(w, []*Task{t})
+				return
+			}
+		}
+
+		// For non-ramp tasks append mode tag.
+		if parallelGPUs {
+			name = fmt.Sprintf("%s · parallel", name)
+		} else {
+			name = fmt.Sprintf("%s · sequential", name)
+		}
+
+		params := taskParams{
+			GPUIndices:        body.GPUIndices,
+			ExcludeGPUIndices: body.ExcludeGPUIndices,
+			SizeMB:            body.SizeMB,
+			BenchmarkProfile:  body.Profile,
+			RunNCCL:           runNCCL,
+			ParallelGPUs:      parallelGPUs,
+			DisplayName:       body.DisplayName,
+		}
+		tasks, err := buildNvidiaTaskSet(target, defaultTaskPriority(target, params), time.Now(), params, name, h.opts.App, "bee-bench-nvidia")
 		if err != nil {
 			writeError(w, http.StatusBadRequest, err.Error())
 			return
@@ -526,56 +702,7 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 }
 
 func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
-	if h.opts.App == nil {
-		writeError(w, http.StatusServiceUnavailable, "app not configured")
-		return
-	}
-
-	var body struct {
-		Profile           string `json:"profile"`
-		SizeMB            int    `json:"size_mb"`
-		GPUIndices        []int  `json:"gpu_indices"`
-		ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
-		RunNCCL           *bool  `json:"run_nccl"`
-		ParallelGPUs      *bool  `json:"parallel_gpus"`
-		DisplayName       string `json:"display_name"`
-	}
-	if r.Body != nil {
-		if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
-			writeError(w, http.StatusBadRequest, "invalid request body")
-			return
-		}
-	}
-
-	runNCCL := true
-	if body.RunNCCL != nil {
-		runNCCL = *body.RunNCCL
-	}
-	parallelGPUs := false
-	if body.ParallelGPUs != nil {
-		parallelGPUs = *body.ParallelGPUs
-	}
-	name := taskDisplayName("nvidia-benchmark", "", "")
-	if strings.TrimSpace(body.DisplayName) != "" {
-		name = body.DisplayName
-	}
-	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
-		GPUIndices:        body.GPUIndices,
-		ExcludeGPUIndices: body.ExcludeGPUIndices,
-		SizeMB:            body.SizeMB,
-		BenchmarkProfile:  body.Profile,
-		RunNCCL:           runNCCL,
-		ParallelGPUs:      parallelGPUs,
-		DisplayName:       body.DisplayName,
-	}, name, h.opts.App, "benchmark-nvidia")
-	if err != nil {
-		writeError(w, http.StatusBadRequest, err.Error())
-		return
-	}
-	for _, t := range tasks {
-		globalQueue.enqueue(t)
-	}
-	writeTaskRunResponse(w, tasks)
+	h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf").ServeHTTP(w, r)
 }
 
 func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
@@ -610,6 +737,9 @@ func (h *handler) handleAPISATAbort(w http.ResponseWriter, r *http.Request) {
 			if t.job != nil {
 				t.job.abort()
 			}
+			if taskMayLeaveOrphanWorkers(t.Target) {
+				platform.KillTestWorkers()
+			}
 			t.Status = TaskCancelled
 			now := time.Now()
 			t.DoneAt = &now
@@ -950,25 +1080,62 @@ func (h *handler) handleAPIRAMStatus(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
 		return
 	}
-	status := h.opts.App.LiveBootSource()
+	status := h.currentRAMStatus()
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(status)
 }
 
+type ramStatusResponse struct {
+	platform.LiveMediaRAMState
+	InstallTaskActive bool   `json:"install_task_active,omitempty"`
+	CopyTaskActive    bool   `json:"copy_task_active,omitempty"`
+	CanStartTask      bool   `json:"can_start_task,omitempty"`
+	BlockedReason     string `json:"blocked_reason,omitempty"`
+}
+
+func (h *handler) currentRAMStatus() ramStatusResponse {
+	state := h.opts.App.LiveMediaRAMState()
+	resp := ramStatusResponse{LiveMediaRAMState: state}
+	if globalQueue.hasActiveTarget("install") {
+		resp.InstallTaskActive = true
+		resp.BlockedReason = "install to disk is already running"
+		return resp
+	}
+	if globalQueue.hasActiveTarget("install-to-ram") {
+		resp.CopyTaskActive = true
+		resp.BlockedReason = "install to RAM task is already pending or running"
+		return resp
+	}
+	if state.InRAM {
+		resp.BlockedReason = "system is already running from RAM"
+		return resp
+	}
+	resp.CanStartTask = state.CanStartCopy
+	if !resp.CanStartTask && resp.BlockedReason == "" {
+		resp.BlockedReason = state.Message
+	}
+	return resp
+}
+
 func (h *handler) handleAPIInstallToRAM(w http.ResponseWriter, r *http.Request) {
 	if h.opts.App == nil {
 		writeError(w, http.StatusServiceUnavailable, "app not configured")
 		return
 	}
-	if globalQueue.hasActiveTarget("install") {
-		writeError(w, http.StatusConflict, "install to disk is already running")
+	status := h.currentRAMStatus()
+	if !status.CanStartTask {
+		msg := strings.TrimSpace(status.BlockedReason)
+		if msg == "" {
+			msg = "install to RAM is not available"
+		}
+		writeError(w, http.StatusConflict, msg)
 		return
 	}
 	t := &Task{
 		ID:        newJobID("install-to-ram"),
 		Name:      "Install to RAM",
 		Target:    "install-to-ram",
-		Priority:  10,
+		Priority:  defaultTaskPriority("install-to-ram", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 	}
@@ -1083,7 +1250,7 @@ func (h *handler) handleAPIInstallRun(w http.ResponseWriter, r *http.Request) {
 		ID:        newJobID("install"),
 		Name:      "Install to Disk",
 		Target:    "install",
-		Priority:  20,
+		Priority:  defaultTaskPriority("install", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 		params: taskParams{
@@ -1359,6 +1526,11 @@ func (h *handler) handleAPINetworkRollback(w http.ResponseWriter, _ *http.Reques
 	writeJSON(w, map[string]string{"status": "rolled back"})
 }
 
+func (h *handler) handleAPIBenchmarkResults(w http.ResponseWriter, r *http.Request) {
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	fmt.Fprint(w, renderBenchmarkResultsCard(h.opts.ExportDir))
+}
+
 func (h *handler) rollbackPendingNetworkChange() error {
 	h.pendingNetMu.Lock()
 	pnc := h.pendingNet
@@ -1375,108 +1547,3 @@ func (h *handler) rollbackPendingNetworkChange() error {
 	}
 	return nil
 }
-
-// ── Display / Screen Resolution ───────────────────────────────────────────────
-
-type displayMode struct {
-	Output  string `json:"output"`
-	Mode    string `json:"mode"`
-	Current bool   `json:"current"`
-}
-
-type displayInfo struct {
-	Output  string        `json:"output"`
-	Modes   []displayMode `json:"modes"`
-	Current string        `json:"current"`
-}
-
-var xrandrOutputRE = regexp.MustCompile(`^(\S+)\s+connected`)
-var xrandrModeRE = regexp.MustCompile(`^\s{3}(\d+x\d+)\s`)
-var xrandrCurrentRE = regexp.MustCompile(`\*`)
-
-func parseXrandrOutput(out string) []displayInfo {
-	var infos []displayInfo
-	var cur *displayInfo
-	for _, line := range strings.Split(out, "\n") {
-		if m := xrandrOutputRE.FindStringSubmatch(line); m != nil {
-			if cur != nil {
-				infos = append(infos, *cur)
-			}
-			cur = &displayInfo{Output: m[1]}
-			continue
-		}
-		if cur == nil {
-			continue
-		}
-		if m := xrandrModeRE.FindStringSubmatch(line); m != nil {
-			isCurrent := xrandrCurrentRE.MatchString(line)
-			mode := displayMode{Output: cur.Output, Mode: m[1], Current: isCurrent}
-			cur.Modes = append(cur.Modes, mode)
-			if isCurrent {
-				cur.Current = m[1]
-			}
-		}
-	}
-	if cur != nil {
-		infos = append(infos, *cur)
-	}
-	return infos
-}
-
-func xrandrCommand(args ...string) *exec.Cmd {
-	cmd := exec.Command("xrandr", args...)
-	env := append([]string{}, os.Environ()...)
-	hasDisplay := false
-	hasXAuthority := false
-	for _, kv := range env {
-		if strings.HasPrefix(kv, "DISPLAY=") && strings.TrimPrefix(kv, "DISPLAY=") != "" {
-			hasDisplay = true
-		}
-		if strings.HasPrefix(kv, "XAUTHORITY=") && strings.TrimPrefix(kv, "XAUTHORITY=") != "" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		env = append(env, "DISPLAY=:0")
-	}
-	if !hasXAuthority {
-		env = append(env, "XAUTHORITY=/home/bee/.Xauthority")
-	}
-	cmd.Env = env
-	return cmd
-}
-
-func (h *handler) handleAPIDisplayResolutions(w http.ResponseWriter, _ *http.Request) {
-	out, err := xrandrCommand().Output()
-	if err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+err.Error())
-		return
-	}
-	writeJSON(w, parseXrandrOutput(string(out)))
-}
-
-func (h *handler) handleAPIDisplaySet(w http.ResponseWriter, r *http.Request) {
-	var req struct {
-		Output string `json:"output"`
-		Mode   string `json:"mode"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&req); err != nil || req.Output == "" || req.Mode == "" {
-		writeError(w, http.StatusBadRequest, "output and mode are required")
-		return
-	}
-	// Validate mode looks like WxH to prevent injection
-	if !regexp.MustCompile(`^\d+x\d+$`).MatchString(req.Mode) {
-		writeError(w, http.StatusBadRequest, "invalid mode format")
-		return
-	}
-	// Validate output name (no special chars)
-	if !regexp.MustCompile(`^[A-Za-z0-9_\-]+$`).MatchString(req.Output) {
-		writeError(w, http.StatusBadRequest, "invalid output name")
-		return
-	}
-	if out, err := xrandrCommand("--output", req.Output, "--mode", req.Mode).CombinedOutput(); err != nil {
-		writeError(w, http.StatusInternalServerError, "xrandr: "+strings.TrimSpace(string(out)))
-		return
-	}
-	writeJSON(w, map[string]string{"status": "ok", "output": req.Output, "mode": req.Mode})
-}

audit/internal/webui/api_test.go

@@ -10,30 +10,6 @@ import (
 	"bee/audit/internal/platform"
 )
 
-func TestXrandrCommandAddsDefaultX11Env(t *testing.T) {
-	t.Setenv("DISPLAY", "")
-	t.Setenv("XAUTHORITY", "")
-
-	cmd := xrandrCommand("--query")
-
-	var hasDisplay bool
-	var hasXAuthority bool
-	for _, kv := range cmd.Env {
-		if kv == "DISPLAY=:0" {
-			hasDisplay = true
-		}
-		if kv == "XAUTHORITY=/home/bee/.Xauthority" {
-			hasXAuthority = true
-		}
-	}
-	if !hasDisplay {
-		t.Fatalf("DISPLAY not injected: %v", cmd.Env)
-	}
-	if !hasXAuthority {
-		t.Fatalf("XAUTHORITY not injected: %v", cmd.Env)
-	}
-}
-
 func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	globalQueue.mu.Lock()
 	originalTasks := globalQueue.tasks
@@ -63,6 +39,9 @@ func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	if got := globalQueue.tasks[0].params.BurnProfile; got != "smoke" {
 		t.Fatalf("burn profile=%q want smoke", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityValidate {
+		t.Fatalf("priority=%d want %d", got, taskPriorityValidate)
+	}
 }
 
 func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
@@ -85,7 +64,7 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
 
 	h := &handler{opts: HandlerOptions{App: &app.App{}}}
-	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/perf/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
 	rec := httptest.NewRecorder()
 
 	h.handleAPIBenchmarkNvidiaRun(rec, req)
@@ -99,8 +78,8 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 		t.Fatalf("tasks=%d want 1", len(globalQueue.tasks))
 	}
 	task := globalQueue.tasks[0]
-	if task.Target != "nvidia-benchmark" {
-		t.Fatalf("target=%q want nvidia-benchmark", task.Target)
+	if task.Target != "nvidia-bench-perf" {
+		t.Fatalf("target=%q want nvidia-bench-perf", task.Target)
 	}
 	if got := task.params.GPUIndices; len(got) != 2 || got[0] != 1 || got[1] != 3 {
 		t.Fatalf("gpu indices=%v want [1 3]", got)
@@ -108,6 +87,9 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	if task.params.RunNCCL {
 		t.Fatal("RunNCCL should reflect explicit false from request")
 	}
+	if task.Priority != taskPriorityBenchmark {
+		t.Fatalf("priority=%d want %d", task.Priority, taskPriorityBenchmark)
+	}
 }
 
 func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
@@ -131,7 +113,7 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
 
 	h := &handler{opts: HandlerOptions{App: &app.App{}}}
-	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/perf/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
 	rec := httptest.NewRecorder()
 
 	h.handleAPIBenchmarkNvidiaRun(rec, req)
@@ -157,6 +139,56 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
 		t.Fatalf("task[1] gpu indices=%v want [2]", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityBenchmark {
+		t.Fatalf("task[0] priority=%d want %d", got, taskPriorityBenchmark)
+	}
+	if got := globalQueue.tasks[1].Priority; got != taskPriorityBenchmark {
+		t.Fatalf("task[1] priority=%d want %d", got, taskPriorityBenchmark)
+	}
+}
+
+func TestHandleAPIBenchmarkPowerFitRampQueuesBenchmarkPowerFitTasks(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H100 PCIe"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/power/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"ramp_up":true}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-power").ServeHTTP(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 3 {
+		t.Fatalf("tasks=%d want 3", len(globalQueue.tasks))
+	}
+	for i, task := range globalQueue.tasks {
+		if task.Target != "nvidia-bench-power" {
+			t.Fatalf("task[%d] target=%q", i, task.Target)
+		}
+		if task.Priority != taskPriorityBenchmark {
+			t.Fatalf("task[%d] priority=%d want %d", i, task.Priority, taskPriorityBenchmark)
+		}
+	}
 }
 
 func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
@@ -199,6 +231,41 @@ func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
 	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
 		t.Fatalf("task[1] gpu indices=%v want [2]", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityValidate {
+		t.Fatalf("task[0] priority=%d want %d", got, taskPriorityValidate)
+	}
+	if got := globalQueue.tasks[1].Priority; got != taskPriorityValidate {
+		t.Fatalf("task[1] priority=%d want %d", got, taskPriorityValidate)
+	}
+}
+
+func TestDefaultTaskPriorityOrder(t *testing.T) {
+	got := []int{
+		defaultTaskPriority("install-to-ram", taskParams{}),
+		defaultTaskPriority("audit", taskParams{}),
+		defaultTaskPriority("cpu", taskParams{}),
+		defaultTaskPriority("cpu", taskParams{StressMode: true}),
+		defaultTaskPriority("nvidia-stress", taskParams{}),
+		defaultTaskPriority("nvidia-bench-perf", taskParams{}),
+		defaultTaskPriority("nvidia-bench-power", taskParams{}),
+	}
+	want := []int{
+		taskPriorityInstallToRAM,
+		taskPriorityAudit,
+		taskPriorityValidate,
+		taskPriorityValidateStress,
+		taskPriorityBurn,
+		taskPriorityBenchmark,
+		taskPriorityBenchmark,
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("priority[%d]=%d want %d", i, got[i], want[i])
+		}
+	}
+	if !(got[0] > got[1] && got[1] > got[2] && got[2] > got[3] && got[3] > got[4] && got[4] > got[5] && got[5] == got[6]) {
+		t.Fatalf("priority order=%v", got)
+	}
 }
 
 func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {

audit/internal/webui/charts_svg.go

@@ -83,6 +83,10 @@ func renderMetricChartSVG(title string, labels []string, times []time.Time, data
 		}
 	}
 
+	// Downsample to at most ~1400 points (one per pixel) before building SVG.
+	times, datasets = downsampleTimeSeries(times, datasets, 1400)
+	pointCount = len(times)
+
 	statsLabel := chartStatsLabel(datasets)
 
 	legendItems := []metricChartSeries{}
@@ -196,6 +200,19 @@ func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, s
 		}
 	}
 
+	// Downsample to at most ~1400 points before building SVG.
+	{
+		datasets := make([][]float64, len(series))
+		for i := range series {
+			datasets[i] = series[i].Values
+		}
+		times, datasets = downsampleTimeSeries(times, datasets, 1400)
+		pointCount = len(times)
+		for i := range series {
+			series[i].Values = datasets[i]
+		}
+	}
+
 	scales := make([]chartScale, len(series))
 	for i := range series {
 		min, max := chartSeriesBounds(series[i].Values)
@@ -626,6 +643,87 @@ func writeTimelineBoundaries(b *strings.Builder, layout chartLayout, start, end
 	b.WriteString(`</g>` + "\n")
 }
 
+// downsampleTimeSeries reduces the time series to at most maxPts points using
+// min-max bucketing. Each bucket contributes the index of its min and max value
+// (using the first full-length dataset as the reference series). All parallel
+// datasets are sampled at those same indices so all series stay aligned.
+// If len(times) <= maxPts the inputs are returned unchanged.
+func downsampleTimeSeries(times []time.Time, datasets [][]float64, maxPts int) ([]time.Time, [][]float64) {
+	n := len(times)
+	if n <= maxPts || maxPts <= 0 {
+		return times, datasets
+	}
+	buckets := maxPts / 2
+	if buckets < 1 {
+		buckets = 1
+	}
+	// Use the first dataset that has the same length as times as the reference
+	// for deciding which two indices to keep per bucket.
+	var ref []float64
+	for _, ds := range datasets {
+		if len(ds) == n {
+			ref = ds
+			break
+		}
+	}
+	selected := make([]int, 0, maxPts)
+	bucketSize := float64(n) / float64(buckets)
+	for b := 0; b < buckets; b++ {
+		lo := int(math.Round(float64(b) * bucketSize))
+		hi := int(math.Round(float64(b+1) * bucketSize))
+		if hi > n {
+			hi = n
+		}
+		if lo >= hi {
+			continue
+		}
+		if ref == nil {
+			selected = append(selected, lo)
+			if hi-1 != lo {
+				selected = append(selected, hi-1)
+			}
+			continue
+		}
+		minIdx, maxIdx := lo, lo
+		for i := lo + 1; i < hi; i++ {
+			if ref[i] < ref[minIdx] {
+				minIdx = i
+			}
+			if ref[i] > ref[maxIdx] {
+				maxIdx = i
+			}
+		}
+		if minIdx <= maxIdx {
+			selected = append(selected, minIdx)
+			if maxIdx != minIdx {
+				selected = append(selected, maxIdx)
+			}
+		} else {
+			selected = append(selected, maxIdx)
+			if minIdx != maxIdx {
+				selected = append(selected, minIdx)
+			}
+		}
+	}
+	outTimes := make([]time.Time, len(selected))
+	for i, idx := range selected {
+		outTimes[i] = times[idx]
+	}
+	outDatasets := make([][]float64, len(datasets))
+	for d, ds := range datasets {
+		if len(ds) != n {
+			outDatasets[d] = ds
+			continue
+		}
+		out := make([]float64, len(selected))
+		for i, idx := range selected {
+			out[i] = ds[idx]
+		}
+		outDatasets[d] = out
+	}
+	return outTimes, outDatasets
+}
+
 func chartXForTime(ts, start, end time.Time, left, right int) float64 {
 	if !end.After(start) {
 		return float64(left+right) / 2

audit/internal/webui/kmsg_watcher.go

@@ -232,7 +232,7 @@ func truncate(s string, max int) string {
 // isSATTarget returns true for task targets that run hardware acceptance tests.
 func isSATTarget(target string) bool {
 	switch target {
-	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
+	case "nvidia", "nvidia-targeted-stress", "nvidia-bench-perf", "nvidia-bench-power", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
 		"nvidia-interconnect", "nvidia-bandwidth", "nvidia-stress", "memory", "memory-stress", "storage",
 		"cpu", "sat-stress", "amd", "amd-mem", "amd-bandwidth", "amd-stress",
 		"platform-stress":

audit/internal/webui/server.go

@@ -261,7 +261,9 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/sat/platform-stress/run", h.handleAPISATRun("platform-stress"))
 	mux.HandleFunc("GET /api/sat/stream", h.handleAPISATStream)
 	mux.HandleFunc("POST /api/sat/abort", h.handleAPISATAbort)
-	mux.HandleFunc("POST /api/benchmark/nvidia/run", h.handleAPIBenchmarkNvidiaRun)
+	mux.HandleFunc("POST /api/bee-bench/nvidia/perf/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf"))
+	mux.HandleFunc("POST /api/bee-bench/nvidia/power/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-power"))
+	mux.HandleFunc("GET /api/benchmark/results", h.handleAPIBenchmarkResults)
 
 	// Tasks
 	mux.HandleFunc("GET /api/tasks", h.handleAPITasksList)
@@ -295,10 +297,6 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	// Tools
 	mux.HandleFunc("GET /api/tools/check", h.handleAPIToolsCheck)
 
-	// Display
-	mux.HandleFunc("GET /api/display/resolutions", h.handleAPIDisplayResolutions)
-	mux.HandleFunc("POST /api/display/set", h.handleAPIDisplaySet)
-
 	// GPU presence / tools
 	mux.HandleFunc("GET /api/gpu/presence", h.handleAPIGPUPresence)
 	mux.HandleFunc("GET /api/gpu/nvidia", h.handleAPIGNVIDIAGPUs)

audit/internal/webui/server_test.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"bee/audit/internal/platform"
+	"bee/audit/internal/schema"
 )
 
 func TestChartLegendNumber(t *testing.T) {
@@ -78,6 +79,16 @@ func TestRecoverMiddlewarePreservesStreamingInterfaces(t *testing.T) {
 	}
 }
 
+func TestBuildRuntimeToRAMRowShowsPartialCopyWarning(t *testing.T) {
+	row := buildRuntimeToRAMRow(schema.RuntimeHealth{ToRAMStatus: "partial"})
+	if row.Status != "WARNING" {
+		t.Fatalf("status=%q want WARNING", row.Status)
+	}
+	if !strings.Contains(row.Issue, "Partial or staged RAM copy detected") {
+		t.Fatalf("issue=%q", row.Issue)
+	}
+}
+
 func TestChartDataFromSamplesUsesFullHistory(t *testing.T) {
 	samples := []platform.LiveMetricSample{
 		{
@@ -637,8 +648,11 @@ func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
 		`href="/benchmark"`,
 		`id="benchmark-gpu-list"`,
 		`/api/gpu/nvidia`,
-		`/api/benchmark/nvidia/run`,
+		`/api/bee-bench/nvidia/perf/run`,
+		`/api/bee-bench/nvidia/power/run`,
 		`benchmark-run-nccl`,
+		`Run Performance Benchmark`,
+		`Run Power / Thermal Fit`,
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("benchmark page missing %q: %s", needle, body)
@@ -649,7 +663,7 @@ func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
 func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	dir := t.TempDir()
 	exportDir := filepath.Join(dir, "export")
-	runDir := filepath.Join(exportDir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	runDir := filepath.Join(exportDir, "bee-bench", "perf", "perf-20260406-120000")
 	if err := os.MkdirAll(runDir, 0755); err != nil {
 		t.Fatal(err)
 	}
@@ -691,10 +705,10 @@ func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	body := rec.Body.String()
 	wantTime := result.GeneratedAt.Local().Format("2006-01-02 15:04:05")
 	for _, needle := range []string{
-		`Benchmark Results`,
+		`Perf Results`,
 		`Composite score by saved benchmark run and GPU.`,
-		`GPU #0 — NVIDIA H100 PCIe`,
-		`GPU #1 — NVIDIA H100 PCIe`,
+		`GPU 0`,
+		`GPU 1`,
 		`#1`,
 		wantTime,
 		`1176.25`,
@@ -1094,6 +1108,7 @@ func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
 	}
 	body := rec.Body.String()
 	for _, needle := range []string{
+		// Runtime Health card — LiveCD checks only
 		`Runtime Health`,
 		`<th>Check</th><th>Status</th><th>Source</th><th>Issue</th>`,
 		`Export Directory`,
@@ -1102,16 +1117,18 @@ func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
 		`CUDA / ROCm`,
 		`Required Utilities`,
 		`Bee Services`,
-		`<td>CPU</td>`,
-		`<td>Memory</td>`,
-		`<td>Storage</td>`,
-		`<td>GPU</td>`,
 		`CUDA runtime is not ready for GPU SAT.`,
 		`Missing: nvidia-smi`,
 		`bee-nvidia=inactive`,
-		`cpu SAT: FAILED`,
-		`storage SAT: FAILED`,
-		`sat:nvidia`,
+		// Hardware Summary card — component health badges
+		`Hardware Summary`,
+		`>CPU<`,
+		`>Memory<`,
+		`>Storage<`,
+		`>GPU<`,
+		`>PSU<`,
+		`badge-warn`, // cpu Warning badge
+		`badge-err`,  // storage Critical badge
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("dashboard missing %q: %s", needle, body)

audit/internal/webui/task_report.go

@@ -233,6 +233,9 @@ func renderTaskReportFragment(report taskReport, charts map[string]string, logTe
 	if benchmarkCard := renderTaskBenchmarkResultsCard(report.Target, logText); benchmarkCard != "" {
 		b.WriteString(benchmarkCard)
 	}
+	if powerCard := renderTaskPowerResultsCard(report.Target, logText); powerCard != "" {
+		b.WriteString(powerCard)
+	}
 
 	if len(report.Charts) > 0 {
 		for _, chart := range report.Charts {
@@ -251,7 +254,9 @@ func renderTaskReportFragment(report taskReport, charts map[string]string, logTe
 }
 
 func renderTaskBenchmarkResultsCard(target, logText string) string {
-	if strings.TrimSpace(target) != "nvidia-benchmark" {
+	switch strings.TrimSpace(target) {
+	case "nvidia-bench-perf":
+	default:
 		return ""
 	}
 	resultPath := taskBenchmarkResultPath(logText)
@@ -263,7 +268,7 @@ func renderTaskBenchmarkResultsCard(target, logText string) string {
 		return ""
 	}
 	return renderBenchmarkResultsCardFromRuns(
-		"Benchmark Results",
+		"Perf Results",
 		"Composite score for this benchmark task.",
 		"No benchmark results were saved for this task.",
 		columns,
@@ -271,15 +276,42 @@ func renderTaskBenchmarkResultsCard(target, logText string) string {
 	)
 }
 
+func renderTaskPowerResultsCard(target, logText string) string {
+	if strings.TrimSpace(target) != "nvidia-bench-power" {
+		return ""
+	}
+	resultPath := taskBenchmarkResultPath(logText)
+	if strings.TrimSpace(resultPath) == "" {
+		return ""
+	}
+	raw, err := os.ReadFile(resultPath)
+	if err != nil {
+		return ""
+	}
+	var result platform.NvidiaPowerBenchResult
+	if err := json.Unmarshal(raw, &result); err != nil {
+		return ""
+	}
+	var b strings.Builder
+	b.WriteString(`<div class="card"><div class="card-head">Power Results</div><div class="card-body">`)
+	if len(result.RecommendedSlotOrder) > 0 {
+		b.WriteString(`<p style="margin-bottom:10px"><strong>Recommended slot order:</strong> ` + html.EscapeString(joinTaskIndices(result.RecommendedSlotOrder)) + `</p>`)
+	}
+	b.WriteString(`<table><tr><th>GPU</th><th>Status</th><th>Max Power</th><th>Applied Limit</th></tr>`)
+	for _, gpu := range result.GPUs {
+		fmt.Fprintf(&b, `<tr><td>GPU %d</td><td>%s</td><td>%.0f W</td><td>%.0f W</td></tr>`,
+			gpu.Index, html.EscapeString(gpu.Status), gpu.MaxObservedPowerW, gpu.AppliedPowerLimitW)
+	}
+	b.WriteString(`</table></div></div>`)
+	return b.String()
+}
+
 func taskBenchmarkResultPath(logText string) string {
 	archivePath := taskArchivePathFromLog(logText)
 	if archivePath == "" {
 		return ""
 	}
 	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
-	if runDir == archivePath {
-		return ""
-	}
 	return filepath.Join(runDir, "result.json")
 }
 

audit/internal/webui/tasks.go

@@ -32,14 +32,14 @@ const (
 var taskNames = map[string]string{
 	"nvidia":                 "NVIDIA SAT",
 	"nvidia-targeted-stress": "NVIDIA Targeted Stress Validate (dcgmi diag targeted_stress)",
-	"nvidia-benchmark":       "NVIDIA Benchmark",
+	"nvidia-bench-perf":      "NVIDIA Bee Bench Perf",
+	"nvidia-bench-power":     "NVIDIA Bee Bench Power",
 	"nvidia-compute":         "NVIDIA Max Compute Load (dcgmproftester)",
 	"nvidia-targeted-power":  "NVIDIA Targeted Power (dcgmi diag targeted_power)",
 	"nvidia-pulse":           "NVIDIA Pulse Test (dcgmi diag pulse_test)",
 	"nvidia-interconnect":    "NVIDIA Interconnect Test (NCCL all_reduce_perf)",
 	"nvidia-bandwidth":       "NVIDIA Bandwidth Test (NVBandwidth)",
 	"nvidia-stress":          "NVIDIA GPU Stress",
-	"hpl":                    "LINPACK (HPL)",
 	"memory":                 "Memory SAT",
 	"storage":                "Storage SAT",
 	"cpu":                    "CPU SAT",
@@ -119,6 +119,7 @@ type taskParams struct {
 	StressMode         bool     `json:"stress_mode,omitempty"`
 	GPUIndices         []int    `json:"gpu_indices,omitempty"`
 	ExcludeGPUIndices  []int    `json:"exclude_gpu_indices,omitempty"`
+	StaggerGPUStart    bool     `json:"stagger_gpu_start,omitempty"`
 	SizeMB             int      `json:"size_mb,omitempty"`
 	Passes             int      `json:"passes,omitempty"`
 	Loader             string   `json:"loader,omitempty"`
@@ -126,6 +127,9 @@ type taskParams struct {
 	BenchmarkProfile   string   `json:"benchmark_profile,omitempty"`
 	RunNCCL            bool     `json:"run_nccl,omitempty"`
 	ParallelGPUs       bool     `json:"parallel_gpus,omitempty"`
+	RampStep           int      `json:"ramp_step,omitempty"`
+	RampTotal          int      `json:"ramp_total,omitempty"`
+	RampRunID          string   `json:"ramp_run_id,omitempty"`
 	DisplayName        string   `json:"display_name,omitempty"`
 	Device             string   `json:"device,omitempty"` // for install
 	PlatformComponents []string `json:"platform_components,omitempty"`
@@ -152,6 +156,38 @@ type burnPreset struct {
 	DurationSec int
 }
 
+type nvidiaRampSpec struct {
+	DurationSec      int
+	StaggerSeconds   int
+	TotalDurationSec int
+}
+
+func resolveMemoryValidatePreset(profile string, stress bool) (sizeMB, passes int) {
+	switch strings.TrimSpace(strings.ToLower(profile)) {
+	case "overnight":
+		return 1024, 2
+	case "acceptance":
+		return 1024, 1
+	case "smoke":
+		return 256, 1
+	}
+	if stress {
+		return 512, 1
+	}
+	return 256, 1
+}
+
+func taskMayLeaveOrphanWorkers(target string) bool {
+	switch strings.TrimSpace(strings.ToLower(target)) {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-targeted-power", "nvidia-pulse",
+		"nvidia-bandwidth", "nvidia-stress", "nvidia-compute", "nvidia-bench-perf",
+		"memory", "memory-stress", "cpu", "sat-stress", "platform-stress":
+		return true
+	default:
+		return false
+	}
+}
+
 func resolveBurnPreset(profile string) burnPreset {
 	switch profile {
 	case "overnight":
@@ -163,6 +199,45 @@ func resolveBurnPreset(profile string) burnPreset {
 	}
 }
 
+func resolveNvidiaRampPlan(profile string, enabled bool, selected []int) (nvidiaRampSpec, error) {
+	base := resolveBurnPreset(profile).DurationSec
+	plan := nvidiaRampSpec{
+		DurationSec:      base,
+		TotalDurationSec: base,
+	}
+	if !enabled {
+		return plan, nil
+	}
+	count := len(selected)
+	if count == 0 {
+		return nvidiaRampSpec{}, fmt.Errorf("staggered NVIDIA burn requires explicit GPU selection")
+	}
+	if count == 1 {
+		return plan, nil
+	}
+
+	switch profile {
+	case "acceptance":
+		plan.StaggerSeconds = 10 * 60
+		plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
+	case "overnight":
+		plan.StaggerSeconds = 60 * 60
+		plan.TotalDurationSec = 8 * 60 * 60
+		minTotal := count * 60 * 60
+		if plan.TotalDurationSec < minTotal {
+			plan.TotalDurationSec = minTotal
+		}
+		if plan.TotalDurationSec > 10*60*60 {
+			return nvidiaRampSpec{}, fmt.Errorf("overnight staggered NVIDIA burn supports at most 10 GPUs")
+		}
+		plan.DurationSec = plan.TotalDurationSec - plan.StaggerSeconds*(count-1)
+	default:
+		plan.StaggerSeconds = 2 * 60
+		plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
+	}
+	return plan, nil
+}
+
 func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions {
 	acceptanceCycles := []platform.PlatformStressCycle{
 		{LoadSec: 85, IdleSec: 5},
@@ -580,7 +655,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = 300
 		}
 		archive, err = a.RunNvidiaTargetedStressValidatePack(ctx, "", dur, t.params.GPUIndices, j.append)
-	case "nvidia-benchmark":
+	case "nvidia-bench-perf":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
@@ -592,6 +667,22 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
 			RunNCCL:           t.params.RunNCCL,
 			ParallelGPUs:      t.params.ParallelGPUs,
+			RampStep:          t.params.RampStep,
+			RampTotal:         t.params.RampTotal,
+			RampRunID:         t.params.RampRunID,
+		}, j.append)
+	case "nvidia-bench-power":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		archive, err = a.RunNvidiaPowerBenchCtx(ctx, app.DefaultBeeBenchPowerDir, platform.NvidiaBenchmarkOptions{
+			Profile:           t.params.BenchmarkProfile,
+			GPUIndices:        t.params.GPUIndices,
+			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+			RampStep:          t.params.RampStep,
+			RampTotal:         t.params.RampTotal,
+			RampRunID:         t.params.RampRunID,
 		}, j.append)
 	case "nvidia-compute":
 		if a == nil {
@@ -602,7 +693,18 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if t.params.BurnProfile != "" && dur <= 0 {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
-		archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, j.append)
+		rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
+		if planErr != nil {
+			err = planErr
+			break
+		}
+		if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
+			dur = rampPlan.DurationSec
+		}
+		if rampPlan.StaggerSeconds > 0 {
+			j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
+		}
+		archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, rampPlan.StaggerSeconds, j.append)
 	case "nvidia-targeted-power":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -652,21 +754,31 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if t.params.BurnProfile != "" && dur <= 0 {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
+		rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
+		if planErr != nil {
+			err = planErr
+			break
+		}
+		if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
+			dur = rampPlan.DurationSec
+		}
+		if rampPlan.StaggerSeconds > 0 {
+			j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
+		}
 		archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
 			DurationSec:       dur,
 			Loader:            t.params.Loader,
 			GPUIndices:        t.params.GPUIndices,
 			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+			StaggerSeconds:    rampPlan.StaggerSeconds,
 		}, j.append)
 	case "memory":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		sizeMB, passes := 256, 1
-		if t.params.StressMode {
-			sizeMB, passes = 1024, 3
-		}
+		sizeMB, passes := resolveMemoryValidatePreset(t.params.BurnProfile, t.params.StressMode)
+		j.append(fmt.Sprintf("Memory validate preset: %d MB x %d pass(es)", sizeMB, passes))
 		archive, err = runMemoryAcceptancePackCtx(a, ctx, "", sizeMB, passes, j.append)
 	case "storage":
 		if a == nil {
@@ -740,19 +852,6 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
 		archive, err = runSATStressPackCtx(a, ctx, "", dur, j.append)
-	case "hpl":
-		if a == nil {
-			err = fmt.Errorf("app not configured")
-			break
-		}
-		opts := platform.HPLOptions{
-			MemFraction: 0.80,
-			NB:          256,
-		}
-		archive, err = func() (string, error) {
-			path, _, runErr := a.RunHPL(ctx, "", opts, j.append)
-			return path, runErr
-		}()
 	case "platform-stress":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -935,6 +1034,9 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 			if t.job != nil {
 				t.job.abort()
 			}
+			if taskMayLeaveOrphanWorkers(t.Target) {
+				platform.KillTestWorkers()
+			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
 			taskSerialEvent(t, "finished with status="+t.Status)
@@ -962,6 +1064,9 @@ func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Reque
 			if t.job != nil {
 				t.job.abort()
 			}
+			if taskMayLeaveOrphanWorkers(t.Target) {
+				platform.KillTestWorkers()
+			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
 			taskSerialEvent(t, "finished with status="+t.Status)
@@ -1066,10 +1171,13 @@ func (q *taskQueue) loadLocked() {
 		q.assignTaskLogPathLocked(t)
 		if t.Status == TaskRunning {
 			// The task was interrupted by a bee-web restart. Child processes
-			// (e.g. bee-gpu-burn-worker) survive the restart in their own
-			// process groups and cannot be cancelled retroactively. Mark the
-			// task as failed so the user can decide whether to re-run it
-			// rather than blindly re-launching duplicate workers.
+			// (e.g. bee-gpu-burn-worker, dcgmi/nvvs) survive the restart in
+			// their own process groups. Kill any matching stale workers before
+			// marking the task failed so the next GPU test does not inherit a
+			// busy DCGM slot or duplicate workers.
+			if taskMayLeaveOrphanWorkers(t.Target) {
+				_ = platform.KillTestWorkers()
+			}
 			now := time.Now()
 			t.Status = TaskFailed
 			t.DoneAt = &now

audit/internal/webui/tasks_test.go

@@ -366,7 +366,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	taskReportMetricsDBPath = metricsPath
 	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
 
-	benchmarkDir := filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	benchmarkDir := filepath.Join(dir, "bee-bench", "perf", "perf-20260406-120000")
 	if err := os.MkdirAll(benchmarkDir, 0755); err != nil {
 		t.Fatal(err)
 	}
@@ -398,14 +398,14 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	}
 	task := &Task{
 		ID:           "task-bench",
-		Name:         "NVIDIA Benchmark",
-		Target:       "nvidia-benchmark",
+		Name:         "NVIDIA Bee Bench Perf",
+		Target:       "nvidia-bench-perf",
 		Status:       TaskDone,
 		CreatedAt:    time.Now().UTC().Add(-time.Minute),
 		ArtifactsDir: artifactsDir,
 	}
 	ensureTaskReportPaths(task)
-	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000.tar.gz") + "\n"
+	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-bench", "perf", "perf-20260406-120000.tar.gz") + "\n"
 	if err := os.WriteFile(task.LogPath, []byte(logText), 0644); err != nil {
 		t.Fatal(err)
 	}
@@ -420,9 +420,9 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	}
 	html := string(body)
 	for _, needle := range []string{
-		`Benchmark Results`,
+		`Perf Results`,
 		`Composite score for this benchmark task.`,
-		`GPU #0 — NVIDIA H100 PCIe`,
+		`GPU 0`,
 		`1176.25`,
 	} {
 		if !strings.Contains(html, needle) {
@@ -491,6 +491,83 @@ func TestResolveBurnPreset(t *testing.T) {
 	}
 }
 
+func TestResolveNvidiaRampPlan(t *testing.T) {
+	tests := []struct {
+		name     string
+		profile  string
+		enabled  bool
+		selected []int
+		want     nvidiaRampSpec
+		wantErr  string
+	}{
+		{
+			name:     "disabled uses base preset",
+			profile:  "acceptance",
+			selected: []int{0, 1},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, TotalDurationSec: 60 * 60},
+		},
+		{
+			name:     "smoke ramp uses two minute steps",
+			profile:  "smoke",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 5 * 60, StaggerSeconds: 2 * 60, TotalDurationSec: 9 * 60},
+		},
+		{
+			name:     "acceptance ramp uses ten minute steps",
+			profile:  "acceptance",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, StaggerSeconds: 10 * 60, TotalDurationSec: 80 * 60},
+		},
+		{
+			name:     "overnight stays at eight hours when possible",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 6 * 60 * 60, StaggerSeconds: 60 * 60, TotalDurationSec: 8 * 60 * 60},
+		},
+		{
+			name:     "overnight extends to keep one hour after final gpu",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2, 3, 4, 5, 6, 7, 8},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, StaggerSeconds: 60 * 60, TotalDurationSec: 9 * 60 * 60},
+		},
+		{
+			name:     "overnight rejects impossible gpu count",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10},
+			wantErr:  "at most 10 GPUs",
+		},
+		{
+			name:    "enabled requires explicit selection",
+			profile: "smoke",
+			enabled: true,
+			wantErr: "requires explicit GPU selection",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := resolveNvidiaRampPlan(tc.profile, tc.enabled, tc.selected)
+			if tc.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.wantErr) {
+					t.Fatalf("err=%v want substring %q", err, tc.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("resolveNvidiaRampPlan error: %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("resolveNvidiaRampPlan(%q, %t, %v)=%+v want %+v", tc.profile, tc.enabled, tc.selected, got, tc.want)
+			}
+		})
+	}
+}
+
 func TestTaskDisplayNameUsesNvidiaStressLoader(t *testing.T) {
 	tests := []struct {
 		loader string
@@ -595,6 +672,36 @@ func TestRunTaskUsesBurnProfileDurationForCPU(t *testing.T) {
 	}
 }
 
+func TestRunTaskUsesQuickPresetForMemoryValidate(t *testing.T) {
+	var gotSizeMB, gotPasses int
+	q := &taskQueue{
+		opts: &HandlerOptions{App: &app.App{}},
+	}
+	tk := &Task{
+		ID:        "mem-validate-1",
+		Name:      "Memory SAT",
+		Target:    "memory",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+		params:    taskParams{StressMode: true},
+	}
+	j := &jobState{}
+
+	orig := runMemoryAcceptancePackCtx
+	runMemoryAcceptancePackCtx = func(_ *app.App, _ context.Context, _ string, sizeMB, passes int, _ func(string)) (string, error) {
+		gotSizeMB = sizeMB
+		gotPasses = passes
+		return "/tmp/memory-validate.tar.gz", nil
+	}
+	defer func() { runMemoryAcceptancePackCtx = orig }()
+
+	q.runTask(tk, j, context.Background())
+
+	if gotSizeMB != 512 || gotPasses != 1 {
+		t.Fatalf("memory validate preset=%dMB x%d want 512MB x1", gotSizeMB, gotPasses)
+	}
+}
+
 func TestRunTaskBuildsSupportBundleWithoutApp(t *testing.T) {
 	dir := t.TempDir()
 	q := &taskQueue{

bible-local/docs/benchmark-clock-calibration.md

@@ -1,5 +1,34 @@
 # Benchmark clock calibration research
 
+## Benchmark methodology versioning
+
+Every benchmark methodology change must bump the benchmark version constant in
+source code by exactly `+1`.
+
+Methodology change means any change that affects comparability of benchmark
+results, including for example:
+- phase durations or phase order
+- enabled/disabled precisions
+- fallback rules
+- normalization rules
+- score formulas or weights
+- degradation thresholds
+- power calibration logic
+- thermal/power penalty logic
+
+Requirements:
+- benchmark version must be stored in source code as an explicit version
+  constant, not inferred from git tag or build metadata
+- benchmark report must always print the benchmark version
+- `result.json` must always include the benchmark version
+- results from different benchmark versions must be treated as non-comparable by
+  default
+
+Purpose:
+- prevent accidental comparison of runs produced by different methodologies
+- make historical benchmark archives self-describing even when detached from git
+- force deliberate version bumps whenever scoring or execution semantics change
+
 ## Status
 In progress. Baseline data from production servers pending.
 

bible-local/docs/gpu-model-propagation.md

@@ -0,0 +1,117 @@
+# GPU Model Name Propagation
+
+How GPU model names are detected, stored, and displayed throughout the project.
+
+---
+
+## Detection Sources
+
+There are **two separate pipelines** for GPU model names — they use different structs and don't share state.
+
+### Pipeline A — Live / SAT (nvidia-smi query at runtime)
+
+**File:** `audit/internal/platform/sat.go`
+
+- `ListNvidiaGPUs()` → `NvidiaGPU.Name` (field: `name`, from `nvidia-smi --query-gpu=index,name,...`)
+- `ListNvidiaGPUStatuses()` → `NvidiaGPUStatus.Name`
+- Used by: GPU selection UI, live metrics labels, burn/stress test logic
+
+### Pipeline B — Benchmark results
+
+**File:** `audit/internal/platform/benchmark.go`, line 124
+
+- `queryBenchmarkGPUInfo(selected)` → `benchmarkGPUInfo.Name`
+- Stored in `BenchmarkGPUResult.Name` (`json:"name,omitempty"`)
+- Used by: benchmark history table, benchmark report
+
+### Pipeline C — Hardware audit JSON (PCIe schema)
+
+**File:** `audit/internal/schema/hardware.go`
+
+- `HardwarePCIeDevice.Model *string` (field name is **Model**, not Name)
+- For AMD GPUs: populated by `audit/internal/collector/amdgpu.go` from `info.Product`
+- For NVIDIA GPUs: **NOT populated** by `audit/internal/collector/nvidia.go` — the NVIDIA enricher sets telemetry/status but skips the Model field
+- Used by: hardware summary page (`hwDescribeGPU` in `pages.go:487`)
+
+---
+
+## Key Inconsistency: NVIDIA PCIe Model is Never Set
+
+`audit/internal/collector/nvidia.go` — `enrichPCIeWithNVIDIAData()` enriches NVIDIA PCIe devices with telemetry and status but does **not** populate `HardwarePCIeDevice.Model`.
+
+This means:
+- Hardware summary page shows "Unknown GPU" for all NVIDIA devices (falls back at `pages.go:486`)
+- AMD GPUs do have their model populated
+
+The fix would be: copy `gpu.Name` from the SAT pipeline into `dev.Model` inside `enrichPCIeWithNVIDIAData`.
+
+---
+
+## Benchmark History "Unknown GPU" Issue
+
+**Symptom:** Benchmark history table shows "GPU #N — Unknown GPU" columns instead of real GPU model names.
+
+**Root cause:** `BenchmarkGPUResult.Name` has tag `json:"name,omitempty"`. If `queryBenchmarkGPUInfo()` fails (warns at `benchmark.go:126`) or returns empty names, the Name field is never set and is omitted from JSON. Loaded results have empty Name → falls back to "Unknown GPU" at `pages.go:2226, 2237`.
+
+This happens for:
+- Older result files saved before the `Name` field was added
+- Runs where nvidia-smi query failed before the benchmark started
+
+---
+
+## Fallback Strings — Current State
+
+| Location | File | Fallback string |
+|---|---|---|
+| Hardware summary (PCIe) | `pages.go:486` | `"Unknown GPU"` |
+| Benchmark report summary | `benchmark_report.go:43` | `"Unknown GPU"` |
+| Benchmark report scorecard | `benchmark_report.go:93` | `"Unknown"` ← inconsistent |
+| Benchmark report detail | `benchmark_report.go:122` | `"Unknown GPU"` |
+| Benchmark history per-GPU col | `pages.go:2226` | `"Unknown GPU"` |
+| Benchmark history parallel col | `pages.go:2237` | `"Unknown GPU"` |
+| SAT status file write | `sat.go:922` | `"unknown"` ← lowercase, inconsistent |
+| GPU selection API | `api.go:163` | `"GPU N"` (no "Unknown") |
+
+**Rule:** all UI fallbacks should use `"Unknown GPU"`. The two outliers are `benchmark_report.go:93` (`"Unknown"`) and `sat.go:922` (`"unknown"`).
+
+---
+
+## GPU Selection UI
+
+**File:** `audit/internal/webui/pages.go`
+
+- Source: `GET /api/gpus` → `api.go` → `ListNvidiaGPUs()` → live nvidia-smi
+- Render: `'GPU ' + gpu.index + ' — ' + gpu.name + ' · ' + mem`
+- Fallback: `gpu.name || 'GPU ' + idx` (JS, line ~1432)
+
+This always shows the correct model because it queries nvidia-smi live. It is **not** connected to benchmark result data.
+
+---
+
+## Data Flow Summary
+
+```
+nvidia-smi (live)
+  └─ ListNvidiaGPUs() → NvidiaGPU.Name
+       ├─ GPU selection UI (always correct)
+       ├─ Live metrics labels (charts_svg.go)
+       └─ SAT/burn status file (sat.go)
+
+nvidia-smi (at benchmark start)
+  └─ queryBenchmarkGPUInfo() → benchmarkGPUInfo.Name
+       └─ BenchmarkGPUResult.Name (json:"name,omitempty")
+            ├─ Benchmark report
+            └─ Benchmark history table columns
+
+nvidia-smi / lspci (audit collection)
+  └─ HardwarePCIeDevice.Model (NVIDIA: NOT populated; AMD: populated)
+       └─ Hardware summary page hwDescribeGPU()
+```
+
+---
+
+## What Needs Fixing
+
+1. **NVIDIA PCIe Model** — `enrichPCIeWithNVIDIAData()` should set `dev.Model = &gpu.Name`
+2. **Fallback consistency** — `benchmark_report.go:93` should say `"Unknown GPU"` not `"Unknown"`; `sat.go:922` should say `"Unknown GPU"` not `"unknown"`
+3. **Old benchmark JSONs** — no fix possible for already-saved results with missing names (display-only issue)

iso/builder/VERSIONS

@@ -1,12 +1,13 @@
 DEBIAN_VERSION=12
 DEBIAN_KERNEL_ABI=auto
 NVIDIA_DRIVER_VERSION=590.48.01
+NVIDIA_FABRICMANAGER_VERSION=590.48.01-1
 NCCL_VERSION=2.28.9-1
 NCCL_CUDA_VERSION=13.0
 NCCL_SHA256=2e6faafd2c19cffc7738d9283976a3200ea9db9895907f337f0c7e5a25563186
 NCCL_TESTS_VERSION=2.13.10
 NVCC_VERSION=12.8
-CUBLAS_VERSION=13.0.2.14-1
+CUBLAS_VERSION=13.1.1.3-1
 CUDA_USERSPACE_VERSION=13.0.96-1
 DCGM_VERSION=4.5.3-1
 JOHN_JUMBO_COMMIT=67fcf9fe5a
@@ -19,7 +20,6 @@ ROCRAND_VERSION=3.2.0.60304-76~22.04
 HIP_RUNTIME_AMD_VERSION=6.3.42134.60304-76~22.04
 HIPBLASLT_VERSION=0.10.0.60304-76~22.04
 COMGR_VERSION=2.8.0.60304-76~22.04
-HPL_VERSION=2.3
-HPL_SHA256=32c5c17d22330e6f2337b681aded51637fb6008d3f0eb7c277b163fadd612830
 GO_VERSION=1.24.0
 AUDIT_VERSION=1.0.0
+MEMTEST_VERSION=6.10-4

iso/builder/auto/config

@@ -23,9 +23,9 @@ lb config noauto \
     --bootloaders "grub-efi,syslinux" \
     --debian-installer none \
     --archive-areas "main contrib non-free non-free-firmware" \
-    --mirror-bootstrap "https://deb.debian.org/debian" \
-    --mirror-chroot "https://deb.debian.org/debian" \
-    --mirror-binary "https://deb.debian.org/debian" \
+    --mirror-bootstrap "http://mirror.mephi.ru/debian/" \
+    --mirror-chroot "http://mirror.mephi.ru/debian/" \
+    --mirror-binary "http://mirror.mephi.ru/debian/" \
     --security true \
     --linux-flavours "amd64" \
     --linux-packages "${LB_LINUX_PACKAGES}" \
@@ -33,6 +33,7 @@ lb config noauto \
     --iso-volume "EASY_BEE_${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
     --iso-application "EASY-BEE-${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
     --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=3 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
+    --debootstrap-options "--include=ca-certificates" \
     --apt-recommends false \
     --chroot-squashfs-compression-type zstd \
     "${@}"

iso/builder/bee-gpu-stress.c

@@ -33,9 +33,10 @@ typedef void *CUstream;
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR 75
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR 76
 #define MAX_STRESS_STREAMS 16
-#define MAX_CUBLAS_PROFILES 5
 #define MIN_PROFILE_BUDGET_BYTES ((size_t)4u * 1024u * 1024u)
 #define MIN_STREAM_BUDGET_BYTES ((size_t)64u * 1024u * 1024u)
+#define MAX_SINGLE_PRECISION_STREAMS 4
+#define MAX_SINGLE_PRECISION_PROFILE_BUDGET_BYTES ((size_t)2u * 1024u * 1024u * 1024u)
 
 static const char *ptx_source =
     ".version 6.0\n"
@@ -297,6 +298,13 @@ static int choose_stream_count(int mp_count, int planned_profiles, size_t total_
     return stream_count;
 }
 
+static size_t clamp_single_precision_profile_budget(size_t profile_budget_bytes) {
+    if (profile_budget_bytes > MAX_SINGLE_PRECISION_PROFILE_BUDGET_BYTES) {
+        return MAX_SINGLE_PRECISION_PROFILE_BUDGET_BYTES;
+    }
+    return profile_budget_bytes;
+}
+
 static void destroy_streams(struct cuda_api *api, CUstream *streams, int count) {
     if (!api->cuStreamDestroy) {
         return;
@@ -643,6 +651,20 @@ static const struct profile_desc k_profiles[] = {
         CUDA_R_16F,
         CUBLAS_COMPUTE_32F_FAST_16F,
     },
+    {
+        "int8_tensor",
+        "int8",
+        75,
+        1,
+        0,
+        0,
+        128,
+        CUDA_R_8I,
+        CUDA_R_8I,
+        CUDA_R_32I,
+        CUDA_R_32I,
+        CUBLAS_COMPUTE_32I,
+    },
     {
         "fp8_e4m3",
         "fp8",
@@ -689,6 +711,21 @@ static const struct profile_desc k_profiles[] = {
 #endif
 };
 
+#define PROFILE_COUNT ((int)(sizeof(k_profiles) / sizeof(k_profiles[0])))
+
+static int profile_allowed_for_run(const struct profile_desc *desc, int cc, const char *precision_filter) {
+    if (!(desc->enabled && cc >= desc->min_cc)) {
+        return 0;
+    }
+    if (precision_filter != NULL) {
+        return strcmp(desc->block_label, precision_filter) == 0;
+    }
+    /* Mixed/all phases intentionally exclude fp64/fp4 for now: both paths are
+     * unstable on the current benchmark fleet and can abort the whole mixed
+     * pass after earlier phases already collected useful telemetry. */
+    return strcmp(desc->block_label, "fp64") != 0 && strcmp(desc->block_label, "fp4") != 0;
+}
+
 static int load_cublaslt(struct cublaslt_api *api) {
     memset(api, 0, sizeof(*api));
     api->lib = dlopen("libcublasLt.so.13", RTLD_NOW | RTLD_LOCAL);
@@ -759,10 +796,12 @@ static int check_cublas(const char *step, cublasStatus_t status) {
 static size_t bytes_for_elements(cudaDataType_t type, uint64_t elements) {
     switch (type) {
         case CUDA_R_32F:
+        case CUDA_R_32I:
             return (size_t)(elements * 4u);
         case CUDA_R_16F:
         case CUDA_R_16BF:
             return (size_t)(elements * 2u);
+        case CUDA_R_8I:
         case CUDA_R_8F_E4M3:
         case CUDA_R_8F_E5M2:
             return (size_t)(elements);
@@ -775,6 +814,16 @@ static size_t bytes_for_elements(cudaDataType_t type, uint64_t elements) {
     }
 }
 
+static cudaDataType_t matmul_scale_type(const struct profile_desc *desc) {
+    if (desc->compute_type == CUBLAS_COMPUTE_32I) {
+        return CUDA_R_32I;
+    }
+    if (desc->compute_type == CUBLAS_COMPUTE_64F) {
+        return CUDA_R_64F;
+    }
+    return CUDA_R_32F;
+}
+
 static size_t fp4_scale_bytes(uint64_t rows, uint64_t cols) {
     uint64_t row_tiles = (rows + 127u) / 128u;
     uint64_t col_tiles = (cols + 63u) / 64u;
@@ -881,11 +930,9 @@ static int prepare_profile(struct cublaslt_api *cublas,
                            CUstream stream,
                            size_t profile_budget_bytes,
                            struct prepared_profile *out) {
-    memset(out, 0, sizeof(*out));
-    out->desc = *desc;
-    out->stream = stream;
-
     size_t bytes_per_cell = 0;
+    size_t attempt_budget = profile_budget_bytes;
+
     bytes_per_cell += bytes_for_elements(desc->a_type, 1);
     bytes_per_cell += bytes_for_elements(desc->b_type, 1);
     bytes_per_cell += bytes_for_elements(desc->c_type, 1);
@@ -894,105 +941,115 @@ static int prepare_profile(struct cublaslt_api *cublas,
         return 0;
     }
 
-    uint64_t dim = choose_square_dim(profile_budget_bytes, bytes_per_cell, desc->min_multiple);
-    out->m = dim;
-    out->n = dim;
-    out->k = dim;
+    while (attempt_budget >= MIN_PROFILE_BUDGET_BYTES) {
+        memset(out, 0, sizeof(*out));
+        out->desc = *desc;
+        out->stream = stream;
 
-    size_t desired_workspace = profile_budget_bytes / 8u;
-    if (desired_workspace > 32u * 1024u * 1024u) {
-        desired_workspace = 32u * 1024u * 1024u;
-    }
-    desired_workspace = round_down_size(desired_workspace, 256u);
+        uint64_t dim = choose_square_dim(attempt_budget, bytes_per_cell, desc->min_multiple);
+        out->m = dim;
+        out->n = dim;
+        out->k = dim;
 
-    size_t a_bytes = 0;
-    size_t b_bytes = 0;
-    size_t c_bytes = 0;
-    size_t d_bytes = 0;
-    size_t scale_bytes = 0;
-    while (1) {
-        a_bytes = bytes_for_elements(desc->a_type, out->k * out->m);
-        b_bytes = bytes_for_elements(desc->b_type, out->k * out->n);
-        c_bytes = bytes_for_elements(desc->c_type, out->m * out->n);
-        d_bytes = bytes_for_elements(desc->d_type, out->m * out->n);
-        scale_bytes = profile_scale_bytes(desc, out->m, out->n, out->k);
+        size_t desired_workspace = attempt_budget / 8u;
+        if (desired_workspace > 32u * 1024u * 1024u) {
+            desired_workspace = 32u * 1024u * 1024u;
+        }
+        desired_workspace = round_down_size(desired_workspace, 256u);
 
-        size_t matrix_bytes = a_bytes + b_bytes + c_bytes + d_bytes + scale_bytes;
-        if (matrix_bytes <= profile_budget_bytes) {
-            size_t remaining = profile_budget_bytes - matrix_bytes;
-            out->workspace_size = desired_workspace;
-            if (out->workspace_size > remaining) {
-                out->workspace_size = round_down_size(remaining, 256u);
+        size_t a_bytes = 0;
+        size_t b_bytes = 0;
+        size_t c_bytes = 0;
+        size_t d_bytes = 0;
+        size_t scale_bytes = 0;
+        while (1) {
+            a_bytes = bytes_for_elements(desc->a_type, out->k * out->m);
+            b_bytes = bytes_for_elements(desc->b_type, out->k * out->n);
+            c_bytes = bytes_for_elements(desc->c_type, out->m * out->n);
+            d_bytes = bytes_for_elements(desc->d_type, out->m * out->n);
+            scale_bytes = profile_scale_bytes(desc, out->m, out->n, out->k);
+
+            size_t matrix_bytes = a_bytes + b_bytes + c_bytes + d_bytes + scale_bytes;
+            if (matrix_bytes <= attempt_budget) {
+                size_t remaining = attempt_budget - matrix_bytes;
+                out->workspace_size = desired_workspace;
+                if (out->workspace_size > remaining) {
+                    out->workspace_size = round_down_size(remaining, 256u);
+                }
+                break;
             }
-            break;
+
+            if (out->m <= (uint64_t)desc->min_multiple) {
+                break;
+            }
+            out->m -= (uint64_t)desc->min_multiple;
+            out->n = out->m;
+            out->k = out->m;
+        }
+        if (out->m < (uint64_t)desc->min_multiple) {
+            attempt_budget /= 2u;
+            continue;
         }
 
-        if (out->m <= (uint64_t)desc->min_multiple) {
-            return 0;
-        }
-        out->m -= (uint64_t)desc->min_multiple;
-        out->n = out->m;
-        out->k = out->m;
-    }
-
-    if (!alloc_filled(cuda, &out->a_dev, a_bytes, 0x11) ||
-        !alloc_filled(cuda, &out->b_dev, b_bytes, 0x11) ||
-        !alloc_filled(cuda, &out->c_dev, c_bytes, 0x00) ||
-        !alloc_filled(cuda, &out->d_dev, d_bytes, 0x00)) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    if (!check_cublas("cublasLtMatmulDescCreate",
-                      cublas->cublasLtMatmulDescCreate(&out->op_desc, desc->compute_type, CUDA_R_32F))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    cublasOperation_t transa = CUBLAS_OP_T;
-    cublasOperation_t transb = CUBLAS_OP_N;
-    if (!check_cublas("set TRANSA",
-                      cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
-                                                             CUBLASLT_MATMUL_DESC_TRANSA,
-                                                             &transa,
-                                                             sizeof(transa))) ||
-        !check_cublas("set TRANSB",
-                      cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
-                                                             CUBLASLT_MATMUL_DESC_TRANSB,
-                                                             &transb,
-                                                             sizeof(transb)))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    if (desc->needs_scalar_scale) {
-        float one = 1.0f;
-        if (!alloc_filled(cuda, &out->a_scale_dev, sizeof(one), 0x00) ||
-            !alloc_filled(cuda, &out->b_scale_dev, sizeof(one), 0x00)) {
+        if (!alloc_filled(cuda, &out->a_dev, a_bytes, 0x11) ||
+            !alloc_filled(cuda, &out->b_dev, b_bytes, 0x11) ||
+            !alloc_filled(cuda, &out->c_dev, c_bytes, 0x00) ||
+            !alloc_filled(cuda, &out->d_dev, d_bytes, 0x00)) {
             destroy_profile(cublas, cuda, out);
             return 0;
         }
-        if (!device_upload(cuda, out->a_scale_dev, &one, sizeof(one)) ||
-            !device_upload(cuda, out->b_scale_dev, &one, sizeof(one))) {
+
+        cudaDataType_t scale_type = matmul_scale_type(desc);
+        if (!check_cublas("cublasLtMatmulDescCreate",
+                          cublas->cublasLtMatmulDescCreate(&out->op_desc, desc->compute_type, scale_type))) {
             destroy_profile(cublas, cuda, out);
             return 0;
         }
-        void *a_scale_ptr = (void *)(uintptr_t)out->a_scale_dev;
-        void *b_scale_ptr = (void *)(uintptr_t)out->b_scale_dev;
-        if (!check_cublas("set A scale ptr",
+
+        cublasOperation_t transa = CUBLAS_OP_T;
+        cublasOperation_t transb = CUBLAS_OP_N;
+        if (!check_cublas("set TRANSA",
                           cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
-                                                                 CUBLASLT_MATMUL_DESC_A_SCALE_POINTER,
-                                                                 &a_scale_ptr,
-                                                                 sizeof(a_scale_ptr))) ||
-            !check_cublas("set B scale ptr",
+                                                                 CUBLASLT_MATMUL_DESC_TRANSA,
+                                                                 &transa,
+                                                                 sizeof(transa))) ||
+            !check_cublas("set TRANSB",
                           cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
-                                                                 CUBLASLT_MATMUL_DESC_B_SCALE_POINTER,
-                                                                 &b_scale_ptr,
-                                                                 sizeof(b_scale_ptr)))) {
+                                                                 CUBLASLT_MATMUL_DESC_TRANSB,
+                                                                 &transb,
+                                                                 sizeof(transb)))) {
             destroy_profile(cublas, cuda, out);
             return 0;
         }
-    }
+
+        if (desc->needs_scalar_scale) {
+            float one = 1.0f;
+            if (!alloc_filled(cuda, &out->a_scale_dev, sizeof(one), 0x00) ||
+                !alloc_filled(cuda, &out->b_scale_dev, sizeof(one), 0x00)) {
+                destroy_profile(cublas, cuda, out);
+                return 0;
+            }
+            if (!device_upload(cuda, out->a_scale_dev, &one, sizeof(one)) ||
+                !device_upload(cuda, out->b_scale_dev, &one, sizeof(one))) {
+                destroy_profile(cublas, cuda, out);
+                return 0;
+            }
+            void *a_scale_ptr = (void *)(uintptr_t)out->a_scale_dev;
+            void *b_scale_ptr = (void *)(uintptr_t)out->b_scale_dev;
+            if (!check_cublas("set A scale ptr",
+                              cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
+                                                                     CUBLASLT_MATMUL_DESC_A_SCALE_POINTER,
+                                                                     &a_scale_ptr,
+                                                                     sizeof(a_scale_ptr))) ||
+                !check_cublas("set B scale ptr",
+                              cublas->cublasLtMatmulDescSetAttribute(out->op_desc,
+                                                                     CUBLASLT_MATMUL_DESC_B_SCALE_POINTER,
+                                                                     &b_scale_ptr,
+                                                                     sizeof(b_scale_ptr)))) {
+                destroy_profile(cublas, cuda, out);
+                return 0;
+            }
+        }
 
 #if defined(CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3)
     if (desc->needs_block_scale) {
@@ -1032,78 +1089,94 @@ static int prepare_profile(struct cublaslt_api *cublas,
     }
 #endif
 
-    if (!check_cublas("create A layout",
-                      cublas->cublasLtMatrixLayoutCreate(&out->a_layout, desc->a_type, out->k, out->m, out->k)) ||
-        !check_cublas("create B layout",
-                      cublas->cublasLtMatrixLayoutCreate(&out->b_layout, desc->b_type, out->k, out->n, out->k)) ||
-        !check_cublas("create C layout",
-                      cublas->cublasLtMatrixLayoutCreate(&out->c_layout, desc->c_type, out->m, out->n, out->m)) ||
-        !check_cublas("create D layout",
-                      cublas->cublasLtMatrixLayoutCreate(&out->d_layout, desc->d_type, out->m, out->n, out->m))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    if (!check_cublas("create preference", cublas->cublasLtMatmulPreferenceCreate(&out->preference))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    if (out->workspace_size > 0) {
-        if (!alloc_filled(cuda, &out->workspace_dev, out->workspace_size, 0x00)) {
+        if (!check_cublas("create A layout",
+                          cublas->cublasLtMatrixLayoutCreate(&out->a_layout, desc->a_type, out->k, out->m, out->k)) ||
+            !check_cublas("create B layout",
+                          cublas->cublasLtMatrixLayoutCreate(&out->b_layout, desc->b_type, out->k, out->n, out->k)) ||
+            !check_cublas("create C layout",
+                          cublas->cublasLtMatrixLayoutCreate(&out->c_layout, desc->c_type, out->m, out->n, out->m)) ||
+            !check_cublas("create D layout",
+                          cublas->cublasLtMatrixLayoutCreate(&out->d_layout, desc->d_type, out->m, out->n, out->m))) {
             destroy_profile(cublas, cuda, out);
             return 0;
         }
+
+        if (!check_cublas("create preference", cublas->cublasLtMatmulPreferenceCreate(&out->preference))) {
+            destroy_profile(cublas, cuda, out);
+            return 0;
+        }
+
+        if (out->workspace_size > 0) {
+            if (!alloc_filled(cuda, &out->workspace_dev, out->workspace_size, 0x00)) {
+                destroy_profile(cublas, cuda, out);
+                return 0;
+            }
+        }
+
+        if (!check_cublas("set workspace",
+                          cublas->cublasLtMatmulPreferenceSetAttribute(
+                              out->preference,
+                              CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+                              &out->workspace_size,
+                              sizeof(out->workspace_size)))) {
+            destroy_profile(cublas, cuda, out);
+            return 0;
+        }
+
+        int found = 0;
+        if (check_cublas("heuristic",
+                         cublas->cublasLtMatmulAlgoGetHeuristic(handle,
+                                                                out->op_desc,
+                                                                out->a_layout,
+                                                                out->b_layout,
+                                                                out->c_layout,
+                                                                out->d_layout,
+                                                                out->preference,
+                                                                1,
+                                                                &out->heuristic,
+                                                                &found)) &&
+            found > 0) {
+            out->ready = 1;
+            return 1;
+        }
+
+        destroy_profile(cublas, cuda, out);
+        attempt_budget = round_down_size(attempt_budget * 3u / 4u, 256u);
+        if (attempt_budget < MIN_PROFILE_BUDGET_BYTES) {
+            break;
+        }
     }
 
-    if (!check_cublas("set workspace",
-                      cublas->cublasLtMatmulPreferenceSetAttribute(
-                          out->preference,
-                          CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
-                          &out->workspace_size,
-                          sizeof(out->workspace_size)))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    int found = 0;
-    if (!check_cublas("heuristic",
-                      cublas->cublasLtMatmulAlgoGetHeuristic(handle,
-                                                             out->op_desc,
-                                                             out->a_layout,
-                                                             out->b_layout,
-                                                             out->c_layout,
-                                                             out->d_layout,
-                                                             out->preference,
-                                                             1,
-                                                             &out->heuristic,
-                                                             &found))) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-    if (found <= 0) {
-        destroy_profile(cublas, cuda, out);
-        return 0;
-    }
-
-    out->ready = 1;
-    return 1;
+    return 0;
 }
 
 static int run_cublas_profile(cublasLtHandle_t handle,
                               struct cublaslt_api *cublas,
                               struct prepared_profile *profile) {
+    int32_t alpha_i32 = 1;
+    int32_t beta_i32 = 0;
+    double alpha_f64 = 1.0;
+    double beta_f64 = 0.0;
     float alpha = 1.0f;
     float beta = 0.0f;
+    const void *alpha_ptr = &alpha;
+    const void *beta_ptr = &beta;
+    if (profile->desc.compute_type == CUBLAS_COMPUTE_32I) {
+        alpha_ptr = &alpha_i32;
+        beta_ptr = &beta_i32;
+    } else if (profile->desc.compute_type == CUBLAS_COMPUTE_64F) {
+        alpha_ptr = &alpha_f64;
+        beta_ptr = &beta_f64;
+    }
     return check_cublas(profile->desc.name,
                         cublas->cublasLtMatmul(handle,
                                                profile->op_desc,
-                                               &alpha,
+                                               alpha_ptr,
                                                (const void *)(uintptr_t)profile->a_dev,
                                                profile->a_layout,
                                                (const void *)(uintptr_t)profile->b_dev,
                                                profile->b_layout,
-                                               &beta,
+                                               beta_ptr,
                                                (const void *)(uintptr_t)profile->c_dev,
                                                profile->c_layout,
                                                (void *)(uintptr_t)profile->d_dev,
@@ -1121,9 +1194,10 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                                int cc_minor,
                                int seconds,
                                int size_mb,
+                               const char *precision_filter,
                                struct stress_report *report) {
     struct cublaslt_api cublas;
-    struct prepared_profile prepared[MAX_STRESS_STREAMS * MAX_CUBLAS_PROFILES];
+    struct prepared_profile prepared[MAX_STRESS_STREAMS * PROFILE_COUNT];
     cublasLtHandle_t handle = NULL;
     CUcontext ctx = NULL;
     CUstream streams[MAX_STRESS_STREAMS] = {0};
@@ -1133,11 +1207,12 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     int active = 0;
     int mp_count = 0;
     int stream_count = 1;
-    int profile_count = (int)(sizeof(k_profiles) / sizeof(k_profiles[0]));
+    int profile_count = PROFILE_COUNT;
     int prepared_count = 0;
     size_t requested_budget = 0;
     size_t total_budget = 0;
     size_t per_profile_budget = 0;
+    int budget_profiles = 0;
 
     memset(report, 0, sizeof(*report));
     snprintf(report->backend, sizeof(report->backend), "cublasLt");
@@ -1158,8 +1233,9 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Count profiles matching the filter (for deciding what to run). */
     for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
-        if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc) {
+        if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
             planned++;
         }
     }
@@ -1170,18 +1246,42 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Count all profiles active on this GPU regardless of filter.
+     * Mixed phases still divide budget across the full precision set, while
+     * single-precision benchmark phases dedicate budget only to active
+     * profiles matching precision_filter. */
+    int planned_total = 0;
+    for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
+        if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
+            planned_total++;
+        }
+    }
+    if (planned_total < planned) {
+        planned_total = planned;
+    }
+    budget_profiles = planned_total;
+    if (precision_filter != NULL) {
+        budget_profiles = planned;
+    }
+    if (budget_profiles <= 0) {
+        budget_profiles = planned_total;
+    }
+
     requested_budget = (size_t)size_mb * 1024u * 1024u;
-    if (requested_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
-        requested_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
+    if (requested_budget < (size_t)budget_profiles * MIN_PROFILE_BUDGET_BYTES) {
+        requested_budget = (size_t)budget_profiles * MIN_PROFILE_BUDGET_BYTES;
     }
     total_budget = clamp_budget_to_free_memory(cuda, requested_budget);
-    if (total_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
-        total_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
+    if (total_budget < (size_t)budget_profiles * MIN_PROFILE_BUDGET_BYTES) {
+        total_budget = (size_t)budget_profiles * MIN_PROFILE_BUDGET_BYTES;
     }
     if (query_multiprocessor_count(cuda, dev, &mp_count) &&
         cuda->cuStreamCreate &&
         cuda->cuStreamDestroy) {
-        stream_count = choose_stream_count(mp_count, planned, total_budget, 1);
+        stream_count = choose_stream_count(mp_count, budget_profiles, total_budget, 1);
+    }
+    if (precision_filter != NULL && stream_count > MAX_SINGLE_PRECISION_STREAMS) {
+        stream_count = MAX_SINGLE_PRECISION_STREAMS;
     }
     if (stream_count > 1) {
         int created = 0;
@@ -1194,18 +1294,22 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         }
     }
     report->stream_count = stream_count;
-    per_profile_budget = total_budget / ((size_t)planned * (size_t)stream_count);
+    per_profile_budget = total_budget / ((size_t)budget_profiles * (size_t)stream_count);
     if (per_profile_budget < MIN_PROFILE_BUDGET_BYTES) {
         per_profile_budget = MIN_PROFILE_BUDGET_BYTES;
     }
+    if (precision_filter != NULL) {
+        per_profile_budget = clamp_single_precision_profile_budget(per_profile_budget);
+    }
     report->buffer_mb = (int)(total_budget / (1024u * 1024u));
     append_detail(report->details,
                   sizeof(report->details),
-                  "requested_mb=%d actual_mb=%d streams=%d mp_count=%d per_worker_mb=%zu\n",
+                  "requested_mb=%d actual_mb=%d streams=%d mp_count=%d budget_profiles=%d per_worker_mb=%zu\n",
                   size_mb,
                   report->buffer_mb,
                   report->stream_count,
                   mp_count,
+                  budget_profiles,
                   per_profile_budget / (1024u * 1024u));
 
     for (int i = 0; i < profile_count; i++) {
@@ -1218,6 +1322,13 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                           desc->min_cc);
             continue;
         }
+        if (!profile_allowed_for_run(desc, cc, precision_filter)) {
+            append_detail(report->details,
+                          sizeof(report->details),
+                          "%s=SKIPPED benchmark_disabled\n",
+                          desc->name);
+            continue;
+        }
         for (int lane = 0; lane < stream_count; lane++) {
             CUstream stream = streams[lane];
             if (prepared_count >= (int)(sizeof(prepared) / sizeof(prepared[0]))) {
@@ -1335,10 +1446,29 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
 }
 #endif
 
+static void print_stress_report(const struct stress_report *report, int device_index, int seconds) {
+    printf("device=%s\n", report->device);
+    printf("device_index=%d\n", device_index);
+    printf("compute_capability=%d.%d\n", report->cc_major, report->cc_minor);
+    printf("backend=%s\n", report->backend);
+    printf("duration_s=%d\n", seconds);
+    printf("buffer_mb=%d\n", report->buffer_mb);
+    printf("streams=%d\n", report->stream_count);
+    printf("iterations=%lu\n", report->iterations);
+    printf("checksum=%llu\n", (unsigned long long)report->checksum);
+    if (report->details[0] != '\0') {
+        printf("%s", report->details);
+    }
+    printf("status=OK\n");
+}
+
 int main(int argc, char **argv) {
     int seconds = 5;
     int size_mb = 64;
     int device_index = 0;
+    const char *precision_filter = NULL; /* NULL = all; else block_label to match */
+    const char *precision_plan = NULL;
+    const char *precision_plan_seconds = NULL;
     for (int i = 1; i < argc; i++) {
         if ((strcmp(argv[i], "--seconds") == 0 || strcmp(argv[i], "-t") == 0) && i + 1 < argc) {
             seconds = atoi(argv[++i]);
@@ -1346,8 +1476,16 @@ int main(int argc, char **argv) {
             size_mb = atoi(argv[++i]);
         } else if ((strcmp(argv[i], "--device") == 0 || strcmp(argv[i], "-d") == 0) && i + 1 < argc) {
             device_index = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--precision") == 0 && i + 1 < argc) {
+            precision_filter = argv[++i];
+        } else if (strcmp(argv[i], "--precision-plan") == 0 && i + 1 < argc) {
+            precision_plan = argv[++i];
+        } else if (strcmp(argv[i], "--precision-plan-seconds") == 0 && i + 1 < argc) {
+            precision_plan_seconds = argv[++i];
         } else {
-            fprintf(stderr, "usage: %s [--seconds N] [--size-mb N] [--device N]\n", argv[0]);
+            fprintf(stderr,
+                    "usage: %s [--seconds N] [--size-mb N] [--device N] [--precision int8|fp8|fp16|fp32|fp64|fp4] [--precision-plan p1,p2,...,mixed] [--precision-plan-seconds s1,s2,...]\n",
+                    argv[0]);
             return 2;
         }
     }
@@ -1407,26 +1545,94 @@ int main(int argc, char **argv) {
     int ok = 0;
 
 #if HAVE_CUBLASLT_HEADERS
-    ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, &report);
+    if (precision_plan != NULL && precision_plan[0] != '\0') {
+        char *plan_copy = strdup(precision_plan);
+        char *plan_seconds_copy = NULL;
+        int phase_seconds[32] = {0};
+        int phase_seconds_count = 0;
+        int phase_ok = 0;
+        if (plan_copy == NULL) {
+            fprintf(stderr, "failed to allocate precision plan buffer\n");
+            return 1;
+        }
+        if (precision_plan_seconds != NULL && precision_plan_seconds[0] != '\0') {
+            plan_seconds_copy = strdup(precision_plan_seconds);
+            if (plan_seconds_copy == NULL) {
+                free(plan_copy);
+                fprintf(stderr, "failed to allocate precision plan seconds buffer\n");
+                return 1;
+            }
+            for (char *sec_token = strtok(plan_seconds_copy, ",");
+                 sec_token != NULL && phase_seconds_count < (int)(sizeof(phase_seconds) / sizeof(phase_seconds[0]));
+                 sec_token = strtok(NULL, ",")) {
+                while (*sec_token == ' ' || *sec_token == '\t') {
+                    sec_token++;
+                }
+                if (*sec_token == '\0') {
+                    continue;
+                }
+                phase_seconds[phase_seconds_count++] = atoi(sec_token);
+            }
+        }
+        int phase_idx = 0;
+        for (char *token = strtok(plan_copy, ","); token != NULL; token = strtok(NULL, ","), phase_idx++) {
+            while (*token == ' ' || *token == '\t') {
+                token++;
+            }
+            if (*token == '\0') {
+                continue;
+            }
+            const char *phase_name = token;
+            const char *phase_filter = token;
+            if (strcmp(token, "mixed") == 0 || strcmp(token, "all") == 0) {
+                phase_filter = NULL;
+            }
+            int phase_duration = seconds;
+            if (phase_idx < phase_seconds_count && phase_seconds[phase_idx] > 0) {
+                phase_duration = phase_seconds[phase_idx];
+            }
+            printf("phase_begin=%s\n", phase_name);
+            fflush(stdout);
+            memset(&report, 0, sizeof(report));
+            ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, phase_duration, size_mb, phase_filter, &report);
+            if (ok) {
+                print_stress_report(&report, device_index, phase_duration);
+                phase_ok = 1;
+            } else {
+                printf("phase_error=%s\n", phase_name);
+                if (report.details[0] != '\0') {
+                    printf("%s", report.details);
+                    if (report.details[strlen(report.details) - 1] != '\n') {
+                        printf("\n");
+                    }
+                }
+                printf("status=FAILED\n");
+            }
+            printf("phase_end=%s\n", phase_name);
+            fflush(stdout);
+        }
+        free(plan_seconds_copy);
+        free(plan_copy);
+        return phase_ok ? 0 : 1;
+    }
+    ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, precision_filter, &report);
 #endif
     if (!ok) {
-        if (!run_ptx_fallback(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, &report)) {
+        if (precision_filter != NULL) {
+            fprintf(stderr,
+                    "requested precision path unavailable: precision=%s device=%s cc=%d.%d\n",
+                    precision_filter,
+                    name,
+                    cc_major,
+                    cc_minor);
+            return 1;
+        }
+        int ptx_mb = size_mb;
+        if (!run_ptx_fallback(&cuda, dev, name, cc_major, cc_minor, seconds, ptx_mb, &report)) {
             return 1;
         }
     }
 
-    printf("device=%s\n", report.device);
-    printf("device_index=%d\n", device_index);
-    printf("compute_capability=%d.%d\n", report.cc_major, report.cc_minor);
-    printf("backend=%s\n", report.backend);
-    printf("duration_s=%d\n", seconds);
-    printf("buffer_mb=%d\n", report.buffer_mb);
-    printf("streams=%d\n", report.stream_count);
-    printf("iterations=%lu\n", report.iterations);
-    printf("checksum=%llu\n", (unsigned long long)report.checksum);
-    if (report.details[0] != '\0') {
-        printf("%s", report.details);
-    }
-    printf("status=OK\n");
+    print_stress_report(&report, device_index, seconds);
     return 0;
 }

iso/builder/build-hpl.sh

@@ -1,331 +0,0 @@
-#!/bin/sh
-# build-hpl.sh — build HPL (High Performance LINPACK) for the bee LiveCD.
-#
-# Downloads HPL 2.3 from netlib, downloads OpenBLAS runtime from the Debian 12
-# apt repo, and compiles xhpl using a minimal single-process MPI stub so that
-# no MPI package is required inside the ISO.
-#
-# The resulting xhpl binary is a standard HPL binary whose output is compatible
-# with the accepted HPL format (WR... Gflops lines).
-#
-# Output:
-#   $CACHE_DIR/bin/xhpl
-#   $CACHE_DIR/lib/libopenblas.so*   (runtime, injected into ISO /usr/lib/)
-
-set -e
-
-HPL_VERSION="$1"
-HPL_SHA256="$2"
-DIST_DIR="$3"
-
-[ -n "$HPL_VERSION" ] || { echo "usage: $0 <hpl-version> <sha256> <dist-dir>"; exit 1; }
-[ -n "$HPL_SHA256"  ] || { echo "usage: $0 <hpl-version> <sha256> <dist-dir>"; exit 1; }
-[ -n "$DIST_DIR"    ] || { echo "usage: $0 <hpl-version> <sha256> <dist-dir>"; exit 1; }
-
-echo "=== HPL ${HPL_VERSION} ==="
-
-CACHE_DIR="${DIST_DIR}/hpl-${HPL_VERSION}"
-CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
-DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/hpl-downloads"
-
-if [ -x "${CACHE_DIR}/bin/xhpl" ]; then
-    echo "=== HPL cached, skipping build ==="
-    echo "binary: ${CACHE_DIR}/bin/xhpl"
-    exit 0
-fi
-
-mkdir -p "${DOWNLOAD_CACHE_DIR}" "${CACHE_DIR}/bin" "${CACHE_DIR}/lib"
-
-# ── download HPL source ────────────────────────────────────────────────────────
-HPL_TAR="${DOWNLOAD_CACHE_DIR}/hpl-${HPL_VERSION}.tar.gz"
-DEFAULT_HPL_URLS="
-https://www.netlib.org/benchmark/hpl/hpl-${HPL_VERSION}.tar.gz
-https://fossies.org/linux/privat/hpl-${HPL_VERSION}.tar.gz
-"
-HPL_GIT_URL="${HPL_GIT_URL:-https://github.com/icl-utk-edu/hpl.git}"
-DEFAULT_HPL_GIT_REFS="v${HPL_VERSION} ${HPL_VERSION} main"
-HPL_SOURCE_MODE="tarball"
-
-download_to_file() {
-    url="$1"
-    out="$2"
-
-    if command -v curl >/dev/null 2>&1; then
-        curl -fL \
-            --connect-timeout 15 \
-            --max-time 180 \
-            --retry 2 \
-            --retry-delay 2 \
-            --output "${out}" \
-            "${url}"
-        return $?
-    fi
-
-    wget \
-        --show-progress \
-        --tries=2 \
-        --timeout=30 \
-        -O "${out}" \
-        "${url}"
-}
-
-download_hpl_tarball() {
-    out="$1"
-    tmp="${out}.part"
-    urls="${HPL_URLS:-$DEFAULT_HPL_URLS}"
-
-    rm -f "${tmp}"
-    for url in ${urls}; do
-        [ -n "${url}" ] || continue
-        echo "=== trying HPL source: ${url} ==="
-        if download_to_file "${url}" "${tmp}"; then
-            mv "${tmp}" "${out}"
-            return 0
-        fi
-        rm -f "${tmp}"
-        echo "=== failed: ${url} ==="
-    done
-
-    echo "ERROR: failed to download HPL ${HPL_VERSION} from all configured URLs" >&2
-    return 1
-}
-
-download_hpl_from_git_archive() {
-    out="$1"
-    refs="${HPL_GIT_REFS:-$DEFAULT_HPL_GIT_REFS}"
-    tmp_root="$(mktemp -d)"
-    repo_dir="${tmp_root}/repo"
-    archive_dir="${tmp_root}/hpl-${HPL_VERSION}"
-    archive_tmp="${out}.part"
-
-    for ref in ${refs}; do
-        [ -n "${ref}" ] || continue
-        echo "=== trying HPL git source: ${HPL_GIT_URL} ref ${ref} ==="
-        rm -rf "${repo_dir}" "${archive_dir}" "${archive_tmp}"
-        if git clone --depth 1 --branch "${ref}" "${HPL_GIT_URL}" "${repo_dir}"; then
-            mv "${repo_dir}" "${archive_dir}"
-            tar czf "${archive_tmp}" -C "${tmp_root}" "hpl-${HPL_VERSION}"
-            mv "${archive_tmp}" "${out}"
-            rm -rf "${tmp_root}"
-            HPL_SOURCE_MODE="git"
-            return 0
-        fi
-        echo "=== failed git ref: ${ref} ==="
-    done
-
-    rm -rf "${tmp_root}" "${archive_tmp}"
-    echo "ERROR: failed to obtain HPL ${HPL_VERSION} from all configured sources" >&2
-    echo "  looked for cache: ${out}" >&2
-    echo "  tarball mirrors: ${HPL_URLS:-$DEFAULT_HPL_URLS}" >&2
-    echo "  git fallback: ${HPL_GIT_URL} refs ${refs}" >&2
-    echo "  override mirrors with HPL_URLS=\"https://mirror1/...\"" >&2
-    echo "  override git refs with HPL_GIT_REFS=\"v${HPL_VERSION} ${HPL_VERSION} main\"" >&2
-    return 1
-}
-
-if [ ! -f "${HPL_TAR}" ]; then
-    echo "=== downloading HPL ${HPL_VERSION} ==="
-    download_hpl_tarball "${HPL_TAR}" || download_hpl_from_git_archive "${HPL_TAR}"
-fi
-
-if [ "${HPL_SOURCE_MODE}" = "tarball" ]; then
-    actual_sha="$(sha256sum "${HPL_TAR}" | awk '{print $1}')"
-    if [ "${actual_sha}" != "${HPL_SHA256}" ]; then
-        echo "ERROR: sha256 mismatch for hpl-${HPL_VERSION}.tar.gz" >&2
-        echo "  expected: ${HPL_SHA256}" >&2
-        echo "  actual:   ${actual_sha}" >&2
-        rm -f "${HPL_TAR}"
-        exit 1
-    fi
-    echo "sha256 OK: hpl-${HPL_VERSION}.tar.gz"
-else
-    echo "=== HPL source obtained from git fallback; skipping tarball sha256 check ==="
-fi
-
-# ── download OpenBLAS from Debian 12 apt repo ─────────────────────────────────
-REPO_BASE="https://deb.debian.org/debian/pool/main/o/openblas"
-PACKAGES_GZ="${DOWNLOAD_CACHE_DIR}/Packages.gz"
-OPENBLAS_PKG="libopenblas0-openmp"
-
-echo "=== fetching Debian 12 Packages.gz ==="
-wget -q -O "${PACKAGES_GZ}" \
-    "https://deb.debian.org/debian/dists/bookworm/main/binary-amd64/Packages.gz"
-
-lookup_deb() {
-    pkg="$1"
-    gzip -dc "${PACKAGES_GZ}" | awk -v pkg="$pkg" '
-        /^Package: / { cur=$2 }
-        /^Filename: / { file=$2 }
-        /^SHA256: /  { sha=$2 }
-        /^$/ {
-            if (cur == pkg) { print file " " sha; exit }
-            cur=""; file=""; sha=""
-        }
-        END {
-            if (cur == pkg) print file " " sha
-        }'
-}
-
-meta="$(lookup_deb "${OPENBLAS_PKG}")"
-[ -n "$meta" ] || { echo "ERROR: ${OPENBLAS_PKG} not found in Packages.gz"; exit 1; }
-repo_file="$(printf '%s' "$meta" | awk '{print $1}')"
-repo_sha="$(printf '%s'  "$meta" | awk '{print $2}')"
-
-OPENBLAS_DEB="${DOWNLOAD_CACHE_DIR}/$(basename "${repo_file}")"
-if [ -f "${OPENBLAS_DEB}" ]; then
-    actual="$(sha256sum "${OPENBLAS_DEB}" | awk '{print $1}')"
-    [ "$actual" = "$repo_sha" ] || rm -f "${OPENBLAS_DEB}"
-fi
-if [ ! -f "${OPENBLAS_DEB}" ]; then
-    echo "=== downloading ${OPENBLAS_PKG} ==="
-    wget --show-progress -O "${OPENBLAS_DEB}" "https://deb.debian.org/debian/${repo_file}"
-    actual="$(sha256sum "${OPENBLAS_DEB}" | awk '{print $1}')"
-    [ "$actual" = "$repo_sha" ] || { echo "ERROR: sha256 mismatch for ${OPENBLAS_PKG}"; rm -f "${OPENBLAS_DEB}"; exit 1; }
-fi
-
-# extract libopenblas shared libs
-TMP_DEB=$(mktemp -d)
-trap 'rm -rf "${TMP_DEB}" "${BUILD_TMP:-}"' EXIT INT TERM
-(
-    cd "${TMP_DEB}"
-    ar x "${OPENBLAS_DEB}"
-    tar xf data.tar.*
-)
-find "${TMP_DEB}" \( -name 'libopenblas*.so*' \) \( -type f -o -type l \) \
-    -exec cp -a {} "${CACHE_DIR}/lib/" \;
-echo "=== OpenBLAS libs: $(ls "${CACHE_DIR}/lib/" | wc -l) files ==="
-
-# also need libopenblas-dev header for compilation (we only need the .so symlink)
-OPENBLAS_SO="$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libopenblas.so.*' -type f | sort | head -1)"
-[ -n "${OPENBLAS_SO}" ] || { echo "ERROR: libopenblas.so not extracted"; exit 1; }
-SONAME="$(basename "${OPENBLAS_SO}")"
-ln -sf "${SONAME}" "${CACHE_DIR}/lib/libopenblas.so" 2>/dev/null || true
-ln -sf "${SONAME}" "${CACHE_DIR}/lib/libblas.so" 2>/dev/null || true
-
-# ── build HPL ─────────────────────────────────────────────────────────────────
-BUILD_TMP=$(mktemp -d)
-
-cd "${BUILD_TMP}"
-tar xf "${HPL_TAR}"
-SRC_DIR="$(find . -maxdepth 1 -type d -name 'hpl-*' | head -1)"
-[ -n "${SRC_DIR}" ] || { echo "ERROR: HPL source dir not found"; exit 1; }
-cd "${SRC_DIR}"
-
-# Write a minimal single-process MPI stub so we don't need an MPI package.
-# HPL only needs these functions for single-process execution.
-cat > "${BUILD_TMP}/mpi_stub.c" <<'MPISTUB'
-#include <stdlib.h>
-#include <string.h>
-#include <sys/time.h>
-
-typedef int MPI_Comm;
-typedef int MPI_Datatype;
-typedef int MPI_Op;
-typedef int MPI_Status;
-typedef int MPI_Request;
-
-#define MPI_COMM_WORLD 0
-#define MPI_SUCCESS    0
-#define MPI_DOUBLE     6
-#define MPI_INT        5
-#define MPI_SUM        0
-#define MPI_MAX        1
-#define MPI_MIN        2
-#define MPI_BYTE       1
-#define MPI_ANY_SOURCE -1
-#define MPI_ANY_TAG    -1
-#define MPI_STATUS_IGNORE ((MPI_Status*)0)
-
-int MPI_Init(int *argc, char ***argv)          { (void)argc; (void)argv; return MPI_SUCCESS; }
-int MPI_Finalize(void)                          { return MPI_SUCCESS; }
-int MPI_Comm_rank(MPI_Comm c, int *rank)        { (void)c; *rank = 0; return MPI_SUCCESS; }
-int MPI_Comm_size(MPI_Comm c, int *size)        { (void)c; *size = 1; return MPI_SUCCESS; }
-int MPI_Bcast(void *b, int n, MPI_Datatype t, int r, MPI_Comm c)
-    { (void)b;(void)n;(void)t;(void)r;(void)c; return MPI_SUCCESS; }
-int MPI_Reduce(const void *s, void *r, int n, MPI_Datatype t, MPI_Op op, int root, MPI_Comm c) {
-    (void)op;(void)root;(void)c;
-    size_t sz = (t==MPI_DOUBLE)?sizeof(double):(t==MPI_INT)?sizeof(int):1;
-    memcpy(r, s, (size_t)n * sz);
-    return MPI_SUCCESS;
-}
-int MPI_Allreduce(const void *s, void *r, int n, MPI_Datatype t, MPI_Op op, MPI_Comm c)
-    { return MPI_Reduce(s,r,n,t,op,0,c); }
-int MPI_Send(const void *b, int n, MPI_Datatype t, int d, int tag, MPI_Comm c)
-    { (void)b;(void)n;(void)t;(void)d;(void)tag;(void)c; return MPI_SUCCESS; }
-int MPI_Recv(void *b, int n, MPI_Datatype t, int s, int tag, MPI_Comm c, MPI_Status *st)
-    { (void)b;(void)n;(void)t;(void)s;(void)tag;(void)c;(void)st; return MPI_SUCCESS; }
-int MPI_Sendrecv(const void *sb, int sn, MPI_Datatype st2, int dest, int stag,
-                 void *rb, int rn, MPI_Datatype rt, int src, int rtag,
-                 MPI_Comm c, MPI_Status *status)
-    { (void)sb;(void)sn;(void)st2;(void)dest;(void)stag;
-      (void)rb;(void)rn;(void)rt;(void)src;(void)rtag;(void)c;(void)status;
-      return MPI_SUCCESS; }
-int MPI_Irecv(void *b, int n, MPI_Datatype t, int s, int tag, MPI_Comm c, MPI_Request *req)
-    { (void)b;(void)n;(void)t;(void)s;(void)tag;(void)c;(void)req; return MPI_SUCCESS; }
-int MPI_Wait(MPI_Request *req, MPI_Status *st)
-    { (void)req;(void)st; return MPI_SUCCESS; }
-int MPI_Abort(MPI_Comm c, int code) { (void)c; exit(code); }
-double MPI_Wtime(void) {
-    struct timeval tv;
-    gettimeofday(&tv, NULL);
-    return (double)tv.tv_sec + (double)tv.tv_usec * 1e-6;
-}
-MPISTUB
-
-# Write Make.bee — HPL makefile configuration
-cat > Make.bee <<MAKEFILE
-SHELL        = /bin/sh
-CD           = cd
-CP           = cp
-LN_S         = ln -s
-MKDIR        = mkdir -p
-RM           = /bin/rm -f
-TOUCH        = touch
-ARCH         = bee
-
-# Directories
-TOPdir       = \$(shell pwd)
-INCdir       = \$(TOPdir)/include
-BINdir       = \$(TOPdir)/bin/\$(ARCH)
-LIBdir       = \$(TOPdir)/lib/\$(ARCH)
-HPLlib       = \$(LIBdir)/libhpl.a
-
-# Compiler
-CC           = gcc
-CCNOOPT      = \$(HPL_DEFS)
-CCFLAGS      = \$(HPL_DEFS) -O3 -march=native -funroll-loops -fomit-frame-pointer
-
-# Linker
-LINKER       = gcc
-LINKFLAGS    = \$(CCFLAGS)
-
-# MPI (single-process stub — no actual MPI needed)
-MPdir        =
-MPinc        = -I${BUILD_TMP}
-MPlib        = ${BUILD_TMP}/mpi_stub.o
-
-# BLAS (OpenBLAS)
-LAdir        = ${CACHE_DIR}/lib
-LAinc        =
-LAlib        = -L\$(LAdir) -Wl,-rpath,/usr/lib -lopenblas
-
-HPL_OPTS     =
-HPL_DEFS     = \$(HPL_OPTS) -DHPL_CALL_CBLAS
-MAKEFILE
-echo "=== Make.bee written ==="
-
-# compile MPI stub
-gcc -O2 -c -o "${BUILD_TMP}/mpi_stub.o" "${BUILD_TMP}/mpi_stub.c"
-
-# build HPL
-echo "=== building HPL ${HPL_VERSION} ==="
-make -j"$(nproc)" arch=bee 2>&1 | tail -20
-
-XHPL_BIN="bin/bee/xhpl"
-[ -x "${XHPL_BIN}" ] || { echo "ERROR: xhpl not found after build"; exit 1; }
-
-cp "${XHPL_BIN}" "${CACHE_DIR}/bin/xhpl"
-chmod +x "${CACHE_DIR}/bin/xhpl"
-echo "=== HPL build complete ==="
-echo "binary: ${CACHE_DIR}/bin/xhpl"
-echo "libs:   $(ls "${CACHE_DIR}/lib/")"

iso/builder/build-in-container.sh

@@ -161,6 +161,7 @@ run_variant() {
             -e GOMODCACHE=/cache/go-mod \
             -e TMPDIR=/cache/tmp \
             -e BEE_CACHE_DIR=/cache/bee \
+            -e BEE_REQUIRE_MEMTEST=1 \
             -w /work \
             "${IMAGE_REF}" \
             sh /work/iso/builder/build.sh --variant "${_v}" \
@@ -175,6 +176,7 @@ run_variant() {
             -e GOMODCACHE=/cache/go-mod \
             -e TMPDIR=/cache/tmp \
             -e BEE_CACHE_DIR=/cache/bee \
+            -e BEE_REQUIRE_MEMTEST=1 \
             -w /work \
             "${IMAGE_REF}" \
             sh /work/iso/builder/build.sh --variant "${_v}"

iso/builder/build.sh

@@ -57,6 +57,7 @@ OVERLAY_STAGE_DIR="${DIST_DIR}/overlay-stage-${BUILD_VARIANT}"
 export BEE_GPU_VENDOR BEE_NVIDIA_MODULE_FLAVOR BUILD_VARIANT
 
 . "${BUILDER_DIR}/VERSIONS"
+export MEMTEST_VERSION
 export PATH="$PATH:/usr/local/go/bin"
 : "${BEE_REQUIRE_MEMTEST:=0}"
 
@@ -775,6 +776,7 @@ run_optional_step_sh() {
         return 0
     fi
 
+    mkdir -p "${LOG_DIR}" 2>/dev/null || true
     step_log="${LOG_DIR}/${step_slug}.log"
     echo ""
     echo "=== optional step: ${step_name} ==="
@@ -798,13 +800,14 @@ start_build_log
 # install them on the fly so NVIDIA modules and ISO kernel always match.
 if [ -z "${DEBIAN_KERNEL_ABI}" ] || [ "${DEBIAN_KERNEL_ABI}" = "auto" ]; then
     echo "=== refreshing apt index to detect current kernel ABI ==="
-    apt-get update -qq
+    apt-get update -qq || echo "WARNING: apt-get update failed, trying cached index"
     DEBIAN_KERNEL_ABI=$(apt-cache depends linux-image-amd64 2>/dev/null \
         | awk '/Depends:.*linux-image-[0-9]/{print $2}' \
         | grep -oE '[0-9]+\.[0-9]+\.[0-9]+-[0-9]+' \
         | head -1)
     if [ -z "${DEBIAN_KERNEL_ABI}" ]; then
         echo "ERROR: could not auto-detect kernel ABI from apt-cache" >&2
+        echo "Hint: set DEBIAN_KERNEL_ABI=x.y.z-N in iso/builder/VERSIONS to skip auto-detection" >&2
         exit 1
     fi
     echo "=== kernel ABI: ${DEBIAN_KERNEL_ABI} ==="
@@ -873,9 +876,37 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
 
     CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
+    echo "=== bee-gpu-burn FP4 header probe ==="
+    fp4_type_match="$(grep -Rsnm 1 'CUDA_R_4F_E2M1' "${CUBLAS_CACHE}/include" 2>/dev/null || true)"
+    fp4_scale_match="$(grep -Rsnm 1 'CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3' "${CUBLAS_CACHE}/include" 2>/dev/null || true)"
+    if [ -n "$fp4_type_match" ]; then
+        echo "fp4_header_symbol=present"
+        echo "$fp4_type_match"
+    else
+        echo "fp4_header_symbol=missing"
+    fi
+    if [ -n "$fp4_scale_match" ]; then
+        echo "fp4_scale_mode_symbol=present"
+        echo "$fp4_scale_match"
+    else
+        echo "fp4_scale_mode_symbol=missing"
+    fi
+
     GPU_STRESS_NEED_BUILD=1
-    if [ -f "$GPU_BURN_WORKER_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_BURN_WORKER_BIN" ]; then
+    if [ -f "$GPU_BURN_WORKER_BIN" ]; then
         GPU_STRESS_NEED_BUILD=0
+        for dep in \
+            "${BUILDER_DIR}/bee-gpu-stress.c" \
+            "${BUILDER_DIR}/VERSIONS"; do
+            if [ "$dep" -nt "$GPU_BURN_WORKER_BIN" ]; then
+                GPU_STRESS_NEED_BUILD=1
+                break
+            fi
+        done
+        if [ "$GPU_STRESS_NEED_BUILD" = "0" ] && \
+            find "${CUBLAS_CACHE}/include" "${CUBLAS_CACHE}/lib" -type f -newer "$GPU_BURN_WORKER_BIN" | grep -q .; then
+            GPU_STRESS_NEED_BUILD=1
+        fi
     fi
 
     if [ "$GPU_STRESS_NEED_BUILD" = "1" ]; then
@@ -889,6 +920,12 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     else
         echo "=== bee-gpu-burn worker up to date, skipping build ==="
     fi
+    echo "=== bee-gpu-burn compiled profile probe ==="
+    if grep -aq 'fp4_e2m1' "$GPU_BURN_WORKER_BIN"; then
+        echo "fp4_profile_string=present"
+    else
+        echo "fp4_profile_string=missing"
+    fi
 fi
 
 echo "=== preparing staged overlay (${BUILD_VARIANT}) ==="
@@ -1148,19 +1185,6 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     echo "=== john injected ==="
 fi
 
-# --- build HPL (CPU LINPACK) — runs on all variants ---
-run_step "build HPL ${HPL_VERSION}" "80-hpl" \
-    sh "${BUILDER_DIR}/build-hpl.sh" "${HPL_VERSION}" "${HPL_SHA256}" "${DIST_DIR}"
-
-HPL_CACHE="${DIST_DIR}/hpl-${HPL_VERSION}"
-mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee"
-cp "${HPL_CACHE}/bin/xhpl" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/xhpl"
-chmod +x "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/xhpl"
-chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-hpl" 2>/dev/null || true
-# Inject OpenBLAS runtime libs needed by xhpl
-cp "${HPL_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/" 2>/dev/null || true
-echo "=== HPL injected: xhpl + $(ls "${HPL_CACHE}/lib/" | wc -l) OpenBLAS libs ==="
-
 # --- embed build metadata ---
 mkdir -p "${OVERLAY_STAGE_DIR}/etc"
 BUILD_DATE="$(date +%Y-%m-%d)"
@@ -1193,7 +1217,6 @@ BUILD_DATE=${BUILD_DATE}
 GIT_COMMIT=${GIT_COMMIT}
 DEBIAN_VERSION=${DEBIAN_VERSION}
 DEBIAN_KERNEL_ABI=${DEBIAN_KERNEL_ABI}
-HPL_VERSION=${HPL_VERSION}
 ${GPU_VERSION_LINE}
 EOF
 
@@ -1239,6 +1262,7 @@ fi
 # --- substitute version placeholders in package list and archive ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     sed -i \
+        -e "s/%%NVIDIA_FABRICMANAGER_VERSION%%/${NVIDIA_FABRICMANAGER_VERSION}/g" \
         -e "s/%%DCGM_VERSION%%/${DCGM_VERSION}/g" \
         "${BUILD_WORK_DIR}/config/package-lists/bee-gpu.list.chroot"
 elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
@@ -1281,7 +1305,7 @@ BEE_GPU_VENDOR_UPPER="$(echo "${BUILD_VARIANT}" | tr 'a-z-' 'A-Z_')"
 export BEE_GPU_VENDOR_UPPER
 
 cd "${LB_DIR}"
-run_step_sh "live-build clean" "80-lb-clean" "lb clean 2>&1 | tail -3"
+run_step_sh "live-build clean" "80-lb-clean" "lb clean --all 2>&1 | tail -3"
 run_step_sh "live-build config" "81-lb-config" "lb config 2>&1 | tail -5"
 dump_memtest_debug "pre-build" "${LB_DIR}"
 run_step_sh "live-build build" "90-lb-build" "lb build 2>&1"

iso/builder/config/bootloaders/grub-pc/grub.cfg

@@ -11,18 +11,18 @@ echo "  Hardware Audit LiveCD"
 echo ""
 
 menuentry "EASY-BEE" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
     initrd  @INITRD_LIVE@
 }
 
 submenu "EASY-BEE (advanced options) -->" {
     menuentry "EASY-BEE — GSP=off" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
         initrd  @INITRD_LIVE@
     }
 
     menuentry "EASY-BEE — KMS (no nomodeset)" {
-        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
         initrd  @INITRD_LIVE@
     }
 

iso/builder/config/bootloaders/grub-pc/theme.cfg

@@ -1,9 +1,9 @@
 set color_normal=light-gray/black
-set color_highlight=white/dark-gray
+set color_highlight=yellow/black
 
 if [ -e /boot/grub/splash.png ]; then
     set theme=/boot/grub/live-theme/theme.txt
 else
-    set menu_color_normal=cyan/black
-    set menu_color_highlight=white/dark-gray
+    set menu_color_normal=yellow/black
+    set menu_color_highlight=white/brown
 fi

iso/builder/config/bootloaders/isolinux/live.cfg.in

@@ -3,31 +3,31 @@ label live-@FLAVOUR@-normal
     menu default
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.nvidia.mode=normal
+    append @APPEND_LIVE@ bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-kms
     menu label EASY-BEE (^graphics/KMS)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-toram
     menu label EASY-BEE (^load to RAM)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ toram bee.nvidia.mode=normal
+    append @APPEND_LIVE@ toram bee.nvidia.mode=normal pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-gsp-off
     menu label EASY-BEE (^NVIDIA GSP=off)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off
+    append @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-kms-gsp-off
     menu label EASY-BEE (g^raphics/KMS, GSP=off)
     linux @LINUX@
     initrd @INITRD@
-    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off
+    append @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1
 
 label live-@FLAVOUR@-failsafe
     menu label EASY-BEE (^fail-safe)

iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot

@@ -25,6 +25,7 @@ ensure_bee_console_user() {
 ensure_bee_console_user
 
 # Enable common bee services
+systemctl enable bee-hpc-tuning.service
 systemctl enable bee-network.service
 systemctl enable bee-preflight.service
 systemctl enable bee-audit.service
@@ -42,6 +43,7 @@ systemctl enable bee-journal-mirror@ttyS1.service 2>/dev/null || true
 # Enable GPU-vendor specific services
 if [ "$GPU_VENDOR" = "nvidia" ]; then
     systemctl enable nvidia-dcgm.service 2>/dev/null || true
+    systemctl enable nvidia-fabricmanager.service 2>/dev/null || true
     systemctl enable bee-nvidia.service
 elif [ "$GPU_VENDOR" = "amd" ]; then
     # ROCm symlinks (packages install to /opt/rocm-*/bin/)
@@ -55,6 +57,7 @@ fi
 # nogpu: no GPU services needed
 
 # Ensure scripts are executable
+chmod +x /usr/local/bin/bee-hpc-tuning  2>/dev/null || true
 chmod +x /usr/local/bin/bee-network.sh  2>/dev/null || true
 chmod +x /usr/local/bin/bee-sshsetup   2>/dev/null || true
 chmod +x /usr/local/bin/bee-smoketest  2>/dev/null || true

iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot

@@ -82,16 +82,22 @@ glow_draw.ellipse((520, 340, 1400, 760), fill=(255, 190, 40, 36))
 glow = glow.filter(ImageFilter.GaussianBlur(60))
 img = Image.alpha_composite(img.convert('RGBA'), glow)
 
-font_logo = load_font(MONO_FONT_CANDIDATES, 64)
+TARGET_LOGO_W = 400
+max_chars = max(len(line) for line in ASCII_ART)
+_probe_font = load_font(MONO_FONT_CANDIDATES, 64)
+_probe_cw, _ = mono_metrics(_probe_font)
+font_size_logo = max(6, int(64 * TARGET_LOGO_W / (_probe_cw * max_chars)))
+font_logo = load_font(MONO_FONT_CANDIDATES, font_size_logo)
 char_w, char_h = mono_metrics(font_logo)
-logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 8)
+logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 2)
 logo_w, logo_h = logo_mask.size
 logo_x = (W - logo_w) // 2
-logo_y = 270
+logo_y = 380
 
-shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(2))
-img.paste(SHADOW, (logo_x + 16, logo_y + 14), shadow_mask)
-img.paste(FG_DIM, (logo_x + 8, logo_y + 7), logo_mask)
+sh_off = max(1, font_size_logo // 6)
+shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(1))
+img.paste(SHADOW, (logo_x + sh_off * 2, logo_y + sh_off * 2), shadow_mask)
+img.paste(FG_DIM, (logo_x + sh_off, logo_y + sh_off), logo_mask)
 img.paste(FG, (logo_x, logo_y), logo_mask)
 
 font_sub = load_font(SUB_FONT_CANDIDATES, 30)

iso/builder/config/hooks/normal/9100-memtest.hook.binary

@@ -5,6 +5,8 @@ set -e
 
 : "${BEE_REQUIRE_MEMTEST:=0}"
 
+# memtest86+ 6.x uses memtest86+.bin (no x64 suffix) for the BIOS binary,
+# while 5.x used memtest86+x64.bin. We normalise both to x64 names in the ISO.
 MEMTEST_FILES="memtest86+x64.bin memtest86+x64.efi"
 BINARY_BOOT_DIR="binary/boot"
 GRUB_CFG="binary/boot/grub/grub.cfg"
@@ -24,15 +26,23 @@ fail_or_warn() {
     return 0
 }
 
+# grub.cfg and live.cfg may not exist yet when binary hooks run — live-build
+# creates them after this hook (lb binary_grub-efi / lb binary_syslinux).
+# The template already has memtest entries hardcoded, so a missing config file
+# here is not an error; validate_iso_memtest() checks the final ISO instead.
+warn_only() {
+    log "WARNING: $1"
+}
+
 copy_memtest_file() {
     src="$1"
-    base="$(basename "$src")"
-    dst="${BINARY_BOOT_DIR}/${base}"
+    dst_name="${2:-$(basename "$src")}"
+    dst="${BINARY_BOOT_DIR}/${dst_name}"
 
     [ -f "$src" ] || return 1
     mkdir -p "${BINARY_BOOT_DIR}"
     cp "$src" "$dst"
-    log "copied ${base} from ${src}"
+    log "copied ${dst_name} from ${src}"
 }
 
 extract_memtest_from_deb() {
@@ -41,14 +51,44 @@ extract_memtest_from_deb() {
 
     log "extracting memtest payload from ${deb}"
     dpkg-deb -x "$deb" "$tmpdir"
-    for f in ${MEMTEST_FILES}; do
-        if [ -f "${tmpdir}/boot/${f}" ]; then
-            copy_memtest_file "${tmpdir}/boot/${f}"
-        fi
-    done
+
+    # EFI binary: both 5.x and 6.x use memtest86+x64.efi
+    if [ -f "${tmpdir}/boot/memtest86+x64.efi" ]; then
+        copy_memtest_file "${tmpdir}/boot/memtest86+x64.efi"
+    fi
+
+    # BIOS binary: 5.x = memtest86+x64.bin, 6.x = memtest86+.bin
+    if [ -f "${tmpdir}/boot/memtest86+x64.bin" ]; then
+        copy_memtest_file "${tmpdir}/boot/memtest86+x64.bin"
+    elif [ -f "${tmpdir}/boot/memtest86+.bin" ]; then
+        copy_memtest_file "${tmpdir}/boot/memtest86+.bin" "memtest86+x64.bin"
+    fi
+
     rm -rf "$tmpdir"
 }
 
+download_and_extract_memtest() {
+    tmpdl="$(mktemp -d)"
+    if [ -n "${MEMTEST_VERSION:-}" ]; then
+        pkg_spec="memtest86+=${MEMTEST_VERSION}"
+    else
+        pkg_spec="memtest86+"
+    fi
+    log "downloading ${pkg_spec} from apt"
+    if ! ( cd "$tmpdl" && apt-get download "$pkg_spec" 2>/dev/null ); then
+        log "apt download failed, retrying after apt-get update"
+        apt-get update -qq >/dev/null 2>&1 || true
+        ( cd "$tmpdl" && apt-get download "$pkg_spec" 2>/dev/null ) || true
+    fi
+    deb="$(find "$tmpdl" -maxdepth 1 -type f -name 'memtest86+*.deb' 2>/dev/null | head -1)"
+    if [ -n "$deb" ]; then
+        extract_memtest_from_deb "$deb"
+    else
+        log "apt download of memtest86+ failed"
+    fi
+    rm -rf "$tmpdl"
+}
+
 ensure_memtest_binaries() {
     missing=0
     for f in ${MEMTEST_FILES}; do
@@ -56,10 +96,15 @@ ensure_memtest_binaries() {
     done
     [ "$missing" -eq 1 ] || return 0
 
+    # 1. Try files already placed by lb binary_memtest or chroot
     for root in chroot/boot /boot; do
         for f in ${MEMTEST_FILES}; do
             [ -f "${BINARY_BOOT_DIR}/${f}" ] || copy_memtest_file "${root}/${f}" || true
         done
+        # 6.x BIOS binary may lack x64 in name — copy with normalised name
+        if [ ! -f "${BINARY_BOOT_DIR}/memtest86+x64.bin" ]; then
+            copy_memtest_file "${root}/memtest86+.bin" "memtest86+x64.bin" || true
+        fi
     done
 
     missing=0
@@ -68,6 +113,7 @@ ensure_memtest_binaries() {
     done
     [ "$missing" -eq 1 ] || return 0
 
+    # 2. Try apt package cache (may be empty if lb binary_memtest already purged)
     for root in cache chroot/var/cache/apt/archives /var/cache/apt/archives; do
         [ -d "$root" ] || continue
         deb="$(find "$root" -type f \( -name 'memtest86+_*.deb' -o -name 'memtest86+*.deb' \) 2>/dev/null | head -1)"
@@ -76,6 +122,15 @@ ensure_memtest_binaries() {
         break
     done
 
+    missing=0
+    for f in ${MEMTEST_FILES}; do
+        [ -f "${BINARY_BOOT_DIR}/${f}" ] || missing=1
+    done
+    [ "$missing" -eq 1 ] || return 0
+
+    # 3. Fallback: download fresh from apt (lb binary_memtest purges the cache)
+    download_and_extract_memtest
+
     missing=0
     for f in ${MEMTEST_FILES}; do
         if [ ! -f "${BINARY_BOOT_DIR}/${f}" ]; then
@@ -88,7 +143,7 @@ ensure_memtest_binaries() {
 
 ensure_grub_entry() {
     [ -f "$GRUB_CFG" ] || {
-        fail_or_warn "missing ${GRUB_CFG}"
+        warn_only "missing ${GRUB_CFG} (will be created by lb binary_grub-efi from template)"
         return 0
     }
 
@@ -114,7 +169,7 @@ EOF
 
 ensure_isolinux_entry() {
     [ -f "$ISOLINUX_CFG" ] || {
-        fail_or_warn "missing ${ISOLINUX_CFG}"
+        warn_only "missing ${ISOLINUX_CFG} (will be created by lb binary_syslinux from template)"
         return 0
     }
 

iso/builder/config/package-lists/bee-nvidia.list.chroot

@@ -5,6 +5,7 @@
 # DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with
 # CUDA 13 userspace, so install the CUDA 13 build plus proprietary components
 # explicitly.
+nvidia-fabricmanager=%%NVIDIA_FABRICMANAGER_VERSION%%
 datacenter-gpu-manager-4-cuda13=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary-cuda13=1:%%DCGM_VERSION%%

iso/overlay/etc/systemd/system/bee-hpc-tuning.service

@@ -0,0 +1,14 @@
+[Unit]
+Description=Bee: HPC tuning (CPU governor, C-states)
+After=local-fs.target
+Before=bee-nvidia.service bee-audit.service
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/bee-log-run /appdata/bee/export/bee-hpc-tuning.log /usr/local/bin/bee-hpc-tuning
+StandardOutput=journal
+StandardError=journal
+RemainAfterExit=yes
+
+[Install]
+WantedBy=multi-user.target

iso/overlay/usr/local/bin/bee-dcgmproftester-staggered

@@ -0,0 +1,110 @@
+#!/bin/sh
+set -eu
+
+SECONDS=300
+STAGGER_SECONDS=180
+DEVICES=""
+EXCLUDE=""
+
+usage() {
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3]" >&2
+    exit 2
+}
+
+normalize_list() {
+    echo "${1:-}" | tr ',' '\n' | sed 's/[[:space:]]//g' | awk 'NF' | sort -n | uniq | paste -sd, -
+}
+
+contains_csv() {
+    needle="$1"
+    haystack="${2:-}"
+    echo ",${haystack}," | grep -q ",${needle},"
+}
+
+resolve_dcgmproftester() {
+    for candidate in dcgmproftester dcgmproftester13 dcgmproftester12 dcgmproftester11; do
+        if command -v "${candidate}" >/dev/null 2>&1; then
+            command -v "${candidate}"
+            return 0
+        fi
+    done
+    return 1
+}
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
+        --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
+        --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        *) usage ;;
+    esac
+done
+
+PROF=$(resolve_dcgmproftester) || { echo "dcgmproftester not found in PATH" >&2; exit 1; }
+ALL_DEVICES=$(nvidia-smi --query-gpu=index --format=csv,noheader,nounits 2>/dev/null | sed 's/[[:space:]]//g' | awk 'NF' | paste -sd, -)
+[ -n "${ALL_DEVICES}" ] || { echo "nvidia-smi found no NVIDIA GPUs" >&2; exit 1; }
+
+DEVICES=$(normalize_list "${DEVICES}")
+EXCLUDE=$(normalize_list "${EXCLUDE}")
+SELECTED="${DEVICES}"
+if [ -z "${SELECTED}" ]; then
+    SELECTED="${ALL_DEVICES}"
+fi
+
+FINAL=""
+for id in $(echo "${SELECTED}" | tr ',' ' '); do
+    [ -n "${id}" ] || continue
+    if contains_csv "${id}" "${EXCLUDE}"; then
+        continue
+    fi
+    if [ -z "${FINAL}" ]; then
+        FINAL="${id}"
+    else
+        FINAL="${FINAL},${id}"
+    fi
+done
+
+[ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
+
+echo "loader=dcgmproftester-staggered"
+echo "selected_gpus=${FINAL}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
+
+TMP_DIR=$(mktemp -d)
+trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
+
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
+gpu_pos=0
+WORKERS=""
+for id in $(echo "${FINAL}" | tr ',' ' '); do
+    gpu_pos=$((gpu_pos + 1))
+    log="${TMP_DIR}/gpu-${id}.log"
+    extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
+    gpu_seconds=$(( SECONDS + extra_sec ))
+    echo "starting gpu ${id} seconds=${gpu_seconds}"
+    CUDA_VISIBLE_DEVICES="${id}" "${PROF}" --no-dcgm-validation -t 1004 -d "${gpu_seconds}" >"${log}" 2>&1 &
+    pid=$!
+    WORKERS="${WORKERS} ${pid}:${id}:${log}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
+done
+
+status=0
+for spec in ${WORKERS}; do
+    pid=${spec%%:*}
+    rest=${spec#*:}
+    id=${rest%%:*}
+    log=${rest#*:}
+    if wait "${pid}"; then
+        echo "gpu ${id} finished: OK"
+    else
+        rc=$?
+        echo "gpu ${id} finished: FAILED rc=${rc}"
+        status=1
+    fi
+    sed "s/^/[gpu ${id}] /" "${log}" || true
+done
+
+exit "${status}"

iso/overlay/usr/local/bin/bee-gpu-burn

@@ -2,13 +2,17 @@
 set -eu
 
 SECONDS=5
+STAGGER_SECONDS=0
 SIZE_MB=0
 DEVICES=""
 EXCLUDE=""
+PRECISION=""
+PRECISION_PLAN=""
+PRECISION_PLAN_SECONDS=""
 WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3] [--precision int8|fp8|fp16|fp32|fp64|fp4] [--precision-plan p1,p2,...,mixed] [--precision-plan-seconds s1,s2,...]" >&2
     exit 2
 }
 
@@ -25,9 +29,13 @@ contains_csv() {
 while [ "$#" -gt 0 ]; do
     case "$1" in
         --seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --size-mb|-m) [ "$#" -ge 2 ] || usage; SIZE_MB="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        --precision) [ "$#" -ge 2 ] || usage; PRECISION="$2"; shift 2 ;;
+        --precision-plan) [ "$#" -ge 2 ] || usage; PRECISION_PLAN="$2"; shift 2 ;;
+        --precision-plan-seconds) [ "$#" -ge 2 ] || usage; PRECISION_PLAN_SECONDS="$2"; shift 2 ;;
         *) usage ;;
     esac
 done
@@ -61,14 +69,18 @@ done
 
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
 
 export CUDA_DEVICE_ORDER="PCI_BUS_ID"
 
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
+GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
+gpu_pos=0
 WORKERS=""
 for id in $(echo "${FINAL}" | tr ',' ' '); do
+    gpu_pos=$((gpu_pos + 1))
     log="${TMP_DIR}/gpu-${id}.log"
     gpu_size_mb="${SIZE_MB}"
     if [ "${gpu_size_mb}" -le 0 ] 2>/dev/null; then
@@ -79,11 +91,22 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
             gpu_size_mb=512
         fi
     fi
-    echo "starting gpu ${id} size=${gpu_size_mb}MB"
+    extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
+    gpu_seconds=$(( SECONDS + extra_sec ))
+    echo "starting gpu ${id} size=${gpu_size_mb}MB seconds=${gpu_seconds}"
+    precision_arg=""
+    [ -n "${PRECISION}" ] && precision_arg="--precision ${PRECISION}"
+    precision_plan_arg=""
+    [ -n "${PRECISION_PLAN}" ] && precision_plan_arg="--precision-plan ${PRECISION_PLAN}"
+    precision_plan_seconds_arg=""
+    [ -n "${PRECISION_PLAN_SECONDS}" ] && precision_plan_seconds_arg="--precision-plan-seconds ${PRECISION_PLAN_SECONDS}"
     CUDA_VISIBLE_DEVICES="${id}" \
-        "${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+        "${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" ${precision_arg} ${precision_plan_arg} ${precision_plan_seconds_arg} >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 
 status=0

iso/overlay/usr/local/bin/bee-hpc-tuning

@@ -0,0 +1,41 @@
+#!/bin/sh
+# bee-hpc-tuning — apply HPC tuning for deterministic benchmarking
+# Called by bee-hpc-tuning.service at boot.
+
+log() { echo "[bee-hpc-tuning] $*"; }
+
+# ── CPU governor ────────────────────────────────────────────────────────────
+# Set all CPU cores to performance governor via sysfs.
+# cpupower is not available; write directly to scaling_governor.
+governor_ok=0
+governor_fail=0
+for gov_path in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
+    [ -f "$gov_path" ] || continue
+    if echo performance > "$gov_path" 2>/dev/null; then
+        governor_ok=$((governor_ok + 1))
+    else
+        governor_fail=$((governor_fail + 1))
+    fi
+done
+
+if [ "$governor_ok" -gt 0 ] && [ "$governor_fail" -eq 0 ]; then
+    log "CPU governor set to performance on ${governor_ok} core(s)"
+elif [ "$governor_ok" -gt 0 ]; then
+    log "WARN: CPU governor: ${governor_ok} OK, ${governor_fail} failed"
+elif [ "$governor_fail" -gt 0 ]; then
+    log "WARN: failed to set CPU governor on ${governor_fail} core(s)"
+else
+    log "WARN: no cpufreq scaling_governor paths found (C-state governor or HW-controlled)"
+fi
+
+# ── Transparent Huge Pages ───────────────────────────────────────────────────
+# Kernel cmdline sets transparent_hugepage=always at boot, but confirm and log.
+thp_path=/sys/kernel/mm/transparent_hugepage/enabled
+if [ -f "$thp_path" ]; then
+    current=$(cat "$thp_path" 2>/dev/null)
+    log "transparent_hugepage: ${current}"
+else
+    log "WARN: transparent_hugepage sysfs path not found"
+fi
+
+log "done"

iso/overlay/usr/local/bin/bee-hpl

@@ -1,97 +0,0 @@
-#!/bin/sh
-# bee-hpl — run HPL (High Performance LINPACK) with auto-sized problem.
-#
-# Generates HPL.dat based on available RAM, runs xhpl, and prints standard
-# HPL output. The WR... line with Gflops is parsed by the bee audit tool.
-#
-# Usage: bee-hpl [--mem-fraction 0.80] [--nb 256] [--seconds N]
-#
-# --mem-fraction   fraction of total RAM to use for the matrix (default 0.80)
-# --nb             block size; 256 is good for modern CPUs (default 256)
-# --seconds        ignored — HPL runtime is determined by problem size; kept
-#                  for interface compatibility with other bee stress tools
-
-set -eu
-
-XHPL="/usr/local/lib/bee/xhpl"
-MEM_FRACTION="0.80"
-NB=256
-
-usage() {
-    echo "usage: $0 [--mem-fraction 0.80] [--nb 256] [--seconds N]" >&2
-    exit 2
-}
-
-while [ "$#" -gt 0 ]; do
-    case "$1" in
-        --mem-fraction) [ "$#" -ge 2 ] || usage; MEM_FRACTION="$2"; shift 2 ;;
-        --nb)           [ "$#" -ge 2 ] || usage; NB="$2"; shift 2 ;;
-        --seconds)      [ "$#" -ge 2 ] || usage; shift 2 ;;  # accepted, ignored
-        *) usage ;;
-    esac
-done
-
-[ -x "${XHPL}" ] || { echo "ERROR: xhpl not found at ${XHPL}" >&2; exit 1; }
-
-# Detect total RAM in bytes
-TOTAL_KB=$(grep MemTotal /proc/meminfo | awk '{print $2}')
-[ -n "${TOTAL_KB}" ] || { echo "ERROR: cannot read MemTotal from /proc/meminfo" >&2; exit 1; }
-TOTAL_BYTES=$(( TOTAL_KB * 1024 ))
-
-# N = floor(sqrt(fraction * total_bytes / 8)) rounded down to multiple of NB
-# Use awk for floating-point sqrt
-N=$(awk -v total="${TOTAL_BYTES}" -v frac="${MEM_FRACTION}" -v nb="${NB}" '
-BEGIN {
-    raw = int(sqrt(total * frac / 8.0))
-    n   = int(raw / nb) * nb
-    if (n < nb) n = nb
-    print n
-}')
-
-echo "loader=bee-hpl"
-echo "total_ram_mb=$(( TOTAL_KB / 1024 ))"
-echo "matrix_n=${N}"
-echo "block_nb=${NB}"
-echo "mem_fraction=${MEM_FRACTION}"
-
-# Generate HPL.dat in a temp directory and run from there
-RUNDIR=$(mktemp -d)
-trap 'rm -rf "${RUNDIR}"' EXIT INT TERM
-
-cat > "${RUNDIR}/HPL.dat" <<DAT
-HPLinpack benchmark input file
-Innovative Computing Laboratory, University of Tennessee
-HPL.out        output file name (if any)
-6              device out (6=stdout, 7=stderr, file)
-1              # of problems sizes (N)
-${N}           Ns
-1              # of NBs
-${NB}          NBs
-0              PMAP process mapping (0=Row-,1=Column-major)
-1              # of process grids (P x Q)
-1              Ps
-1              Qs
-16.0           threshold
-1              # of panel fact
-2              PFACTs (0=left, 1=Crout, 2=Right)
-1              # of recursive stopping criterium
-4              NBMINs (>= 1)
-1              # of panels in recursion
-2              NDIVs
-1              # of recursive panel fact.
-1              RFACTs (0=left, 1=Crout, 2=Right)
-1              # of broadcast
-1              BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
-1              # of lookahead depth
-1              DEPTHs (>=0)
-2              SWAP (0=bin-exch,1=long,2=mix)
-64             swapping threshold
-0              L1 in (0=transposed,1=no-transposed) form
-0              U  in (0=transposed,1=no-transposed) form
-1              Equilibration (0=no,1=yes)
-8              memory alignment in double (> 0)
-DAT
-
-cd "${RUNDIR}"
-echo "---"
-"${XHPL}"

iso/overlay/usr/local/bin/bee-john-gpu-stress

@@ -2,6 +2,7 @@
 set -eu
 
 DURATION_SEC=300
+STAGGER_SECONDS=0
 DEVICES=""
 EXCLUDE=""
 FORMAT=""
@@ -12,7 +13,7 @@ export OCL_ICD_VENDORS="/etc/OpenCL/vendors"
 export LD_LIBRARY_PATH="/usr/lib:/usr/local/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
     exit 2
 }
 
@@ -118,6 +119,7 @@ ensure_opencl_ready() {
 while [ "$#" -gt 0 ]; do
     case "$1" in
         --seconds|-t) [ "$#" -ge 2 ] || usage; DURATION_SEC="$2"; shift 2 ;;
+        --stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
         --format) [ "$#" -ge 2 ] || usage; FORMAT="$2"; shift 2 ;;
@@ -170,6 +172,7 @@ done
 echo "loader=john"
 echo "selected_gpus=${FINAL}"
 echo "john_devices=${JOHN_DEVICES}"
+echo "stagger_seconds=${STAGGER_SECONDS}"
 
 cd "${JOHN_DIR}"
 
@@ -232,14 +235,21 @@ trap cleanup EXIT INT TERM
 echo "format=${CHOSEN_FORMAT}"
 echo "target_seconds=${DURATION_SEC}"
 echo "slice_seconds=${TEST_SLICE_SECONDS}"
-DEADLINE=$(( $(date +%s) + DURATION_SEC ))
+TOTAL_DEVICES=$(echo "${JOHN_DEVICES}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
 _first=1
+pos=0
 for opencl_id in $(echo "${JOHN_DEVICES}" | tr ',' ' '); do
+    pos=$((pos + 1))
     [ "${_first}" = "1" ] || sleep 3
     _first=0
-    run_john_loop "${opencl_id}" "${DEADLINE}" &
+    extra_sec=$(( STAGGER_SECONDS * (TOTAL_DEVICES - pos) ))
+    deadline=$(( $(date +%s) + DURATION_SEC + extra_sec ))
+    run_john_loop "${opencl_id}" "${deadline}" &
     pid=$!
     PIDS="${PIDS} ${pid}"
+    if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${pos}" -lt "${TOTAL_DEVICES}" ]; then
+        sleep "${STAGGER_SECONDS}"
+    fi
 done
 FAIL=0
 for pid in ${PIDS}; do

iso/overlay/usr/local/bin/bee-nvidia-load

@@ -21,8 +21,13 @@ read_nvidia_modules_flavor() {
 
 log "kernel: $(uname -r)"
 
-# Skip if no NVIDIA GPU present (PCI vendor 10de)
-if ! lspci -nn 2>/dev/null | grep -qi '10de:'; then
+# Skip if no NVIDIA display/compute GPU is present.
+# Match only display-class PCI functions (0300 VGA, 0302 3D controller) from vendor 10de.
+have_nvidia_gpu() {
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
+}
+
+if ! have_nvidia_gpu; then
     log "no NVIDIA GPU detected — skipping module load"
     exit 0
 fi
@@ -253,6 +258,22 @@ else
     log "WARN: nvidia-smi not found — cannot enable persistence mode"
 fi
 
+# Start or refresh Fabric Manager after the NVIDIA stack is ready. On NVSwitch
+# systems CUDA/DCGM can report "system not yet initialized" until fabric
+# training completes under nvidia-fabricmanager.
+if command -v systemctl >/dev/null 2>&1 && systemctl list-unit-files --no-legend 2>/dev/null | grep -q '^nvidia-fabricmanager\.service'; then
+    if systemctl restart nvidia-fabricmanager.service >/dev/null 2>&1; then
+        log "nvidia-fabricmanager restarted"
+    elif systemctl start nvidia-fabricmanager.service >/dev/null 2>&1; then
+        log "nvidia-fabricmanager started"
+    else
+        log "WARN: failed to start nvidia-fabricmanager.service"
+        systemctl status nvidia-fabricmanager.service --no-pager 2>&1 | sed 's/^/  fabricmanager: /' || true
+    fi
+else
+    log "WARN: nvidia-fabricmanager.service not installed"
+fi
+
 # Start DCGM host engine so dcgmi can discover GPUs.
 # nv-hostengine must run after the NVIDIA modules and device nodes are ready.
 # If it started too early (for example via systemd before bee-nvidia-load), it can

iso/overlay/usr/local/bin/bee-openbox-session

@@ -9,9 +9,9 @@ xset s noblank
 
 # Set desktop background.
 if [ -f /usr/share/bee/wallpaper.png ]; then
-    feh --bg-fill /usr/share/bee/wallpaper.png
+    feh --bg-center --image-bg '#000000' /usr/share/bee/wallpaper.png
 else
-    xsetroot -solid '#f6c90e'
+    xsetroot -solid '#000000'
 fi
 
 tint2 &

iso/overlay/usr/local/bin/bee-selfheal

@@ -14,7 +14,7 @@ log() {
 }
 
 have_nvidia_gpu() {
-    lspci -nn 2>/dev/null | grep -qi '10de:'
+    lspci -Dn 2>/dev/null | awk '$2 ~ /^03(00|02):$/ && $3 ~ /^10de:/ { found=1; exit } END { exit(found ? 0 : 1) }'
 }
 
 service_active() {