Fix power calibration failing due to DCGM resource contention

When a targeted_power attempt is cancelled (e.g. after sw_thermal
throttle), nv-hostengine holds the diagnostic slot asynchronously.
The next attempt immediately received DCGM_ST_IN_USE (exit 222)
and incorrectly derated the power limit.

Now: exit 222 is detected via isDCGMResourceBusy and triggers an
exponential back-off retry at the same power limit (1s, 2s, 4s, …
up to 256s). Once the back-off delay would exceed 300s the
calibration fails, indicating the slot is persistently held.

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

.gitignore

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

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)
 }
@@ -117,6 +124,7 @@ 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)
+	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)
@@ -562,11 +570,18 @@ 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) 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

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)

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 +48,19 @@ 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)
+	}
 	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
 	b.WriteString("\n")
 
@@ -83,10 +81,28 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		b.WriteString("\n")
 	}
 
+	// ── Methodology ───────────────────────────────────────────────────────────
+	b.WriteString("## Methodology\n\n")
+	fmt.Fprintf(&b, "- Profile `%s` uses standardized baseline -> warmup -> steady-state -> interconnect phases.\n", result.BenchmarkProfile)
+	b.WriteString("- Single-GPU compute score comes from `bee-gpu-burn` on the cuBLASLt path when available.\n")
+	b.WriteString("- Thermal and power limits are inferred from NVIDIA clock-event counters plus sustained telemetry.\n")
+	b.WriteString("- `result.json` is the canonical machine-readable source for the run.\n\n")
+	b.WriteString("**Compute score** is derived from two phases:\n\n")
+	b.WriteString("- **Synthetic** — each precision type (int8, fp8, fp16, fp32, fp64, fp4) runs alone for a dedicated window. ")
+	b.WriteString("Measures peak throughput with the full GPU dedicated to one kernel type. ")
+	b.WriteString("Each result is normalised to fp32-equivalent TOPS using precision weights: ")
+	b.WriteString("fp64 ×2.0 · fp32 ×1.0 · fp16 ×0.5 · int8 ×0.25 · fp8 ×0.25 · fp4 ×0.125.\n")
+	b.WriteString("- **Mixed** — all precision types run simultaneously (combined phase). ")
+	b.WriteString("Reflects real inference workloads where fp8 matrix ops, fp16 attention and fp32 accumulation compete for bandwidth and SM scheduler slots.\n\n")
+	b.WriteString("**Formula:** `Compute = Synthetic × (1 + MixedEfficiency × 0.3)`\n\n")
+	b.WriteString("where `MixedEfficiency = Mixed / Synthetic`. A GPU that sustains 90 % throughput under mixed load ")
+	b.WriteString("receives a +27 % bonus over its synthetic score; one that drops to 60 % receives +18 %.\n\n")
+	b.WriteString("**Composite score** = `Compute × quality_factor` where quality factors in power sustain, thermal sustain, stability, and interconnect.\n\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")
+	b.WriteString("| GPU | Status | Composite | Compute | Synthetic | Mixed | Mixed Eff. | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
+	b.WriteString("|-----|--------|-----------|---------|-----------|-------|------------|-------------|---------------|-----------------|-----------|-------------|\n")
 	for _, gpu := range result.GPUs {
 		name := strings.TrimSpace(gpu.Name)
 		if name == "" {
@@ -100,11 +116,26 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		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",
+		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)
+		}
+		fmt.Fprintf(&b, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %s | %s | %s | %.1f | %.1f | %.1f | %s |\n",
 			gpu.Index, name,
 			gpu.Status,
 			gpu.Scores.CompositeScore,
 			gpu.Scores.ComputeScore,
+			synthetic,
+			mixed,
+			mixedEff,
 			topsPerSM,
 			gpu.Scores.PowerSustainScore,
 			gpu.Scores.ThermalSustainScore,
@@ -139,6 +170,16 @@ 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 after %d targeted_power attempt(s)\n", gpu.PowerCalibrationTries)
+		}
+		if gpu.CalibratedPeakPowerW > 0 {
+			if gpu.CalibratedPeakTempC > 0 {
+				fmt.Fprintf(&b, "- **Power calibration (`dcgmi targeted_power`):** %.0f W p95 at %.1f °C p95\n", gpu.CalibratedPeakPowerW, gpu.CalibratedPeakTempC)
+			} else {
+				fmt.Fprintf(&b, "- **Power calibration (`dcgmi targeted_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)
 		}
@@ -154,6 +195,38 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
 		b.WriteString("\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)
 		if throttle != "none" {
@@ -163,12 +236,14 @@ 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 +304,41 @@ 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")
-
 	// ── 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 +378,7 @@ 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) != 7 || len(phases) != 7 {
+		t.Fatalf("labels=%d phases=%d want 7", 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) != "60,60,60,60,60,60,300" {
+		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) != "300,300,300,300,300,300,3600" {
+		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) != "3600,3600,3600,3600,3600,3600,14400" {
+		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,fp64" {
+		t.Fatalf("supported=%v", got)
+	}
+	if got := benchmarkSupportedPrecisions("10.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32,fp64,fp4" {
+		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)
@@ -147,36 +291,27 @@ 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 TestEnrichGPUInfoWithMaxClocks(t *testing.T) {

audit/internal/platform/benchmark_types.go

@@ -2,6 +2,40 @@ 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"`
+	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"
@@ -14,10 +48,12 @@ 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"`
@@ -25,11 +61,18 @@ type NvidiaBenchmarkResult struct {
 	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"`
 	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"`
@@ -52,30 +95,42 @@ 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"`
+	// 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"`
 }
 
 type BenchmarkTelemetrySummary struct {
@@ -105,6 +160,18 @@ 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 +182,31 @@ 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"`
 	// 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 +224,22 @@ 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 +251,45 @@ 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"`
+	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  float64  `json:"applied_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"`
+	OccupiedSlots       []int    `json:"occupied_slots,omitempty"`
+	OccupiedSlotsNote   string   `json:"occupied_slots_note,omitempty"`
+	Notes               []string `json:"notes,omitempty"`
+}
+
+type NvidiaPowerBenchStep struct {
+	StepIndex              int      `json:"step_index"`
+	GPUIndices             []int    `json:"gpu_indices"`
+	TotalObservedPowerW    float64  `json:"total_observed_power_w,omitempty"`
+	AvgObservedPowerW      float64  `json:"avg_observed_power_w,omitempty"`
+	MinPowerRealizationPct float64  `json:"min_power_realization_pct,omitempty"`
+	AvgPowerRealizationPct float64  `json:"avg_power_realization_pct,omitempty"`
+	DeratedGPUCount        int      `json:"derated_gpu_count,omitempty"`
+	Status                 string   `json:"status"`
+	Notes                  []string `json:"notes,omitempty"`
+}

audit/internal/platform/gpu_metrics.go

@@ -13,14 +13,20 @@ 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"`
 }
 
 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
@@ -141,14 +147,24 @@ 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\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
+		}
+		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%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)
 	}
 	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 +179,25 @@ 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 +207,39 @@ 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 +334,23 @@ 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",
@@ -382,221 +449,6 @@ func drawGPUChartSVG(rows []GPUMetricRow, gpuIdx int) string {
 	return b.String()
 }
 
-const (
-	ansiAmber  = "\033[38;5;214m"
-	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)", ansiAmber, func(r GPUMetricRow) float64 { return r.TempC }},
-		{"GPU Usage (%)", ansiAmber, func(r GPUMetricRow) float64 { return r.UsagePct }},
-		{"Power (W)", ansiAmber, func(r GPUMetricRow) float64 { return r.PowerW }},
-		{"Clock (MHz)", ansiAmber, 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
@@ -641,3 +493,57 @@ 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/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/runtime.go

@@ -171,25 +171,28 @@ func resolvedToolStatus(display string, candidates ...string) ToolStatus {
 	return ToolStatus{Name: display}
 }
 
-// collectToRAMHealth checks whether the LiveCD ISO has been copied to RAM.
-// Status values: "ok" = in RAM, "warning" = toram not active (no copy attempted),
-// "failed" = toram was requested but medium is not in RAM (copy failed or in progress).
+// 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) {
-	inRAM := s.IsLiveMediaInRAM()
-	active := toramActive()
-	switch {
-	case inRAM:
-		health.ToRAMStatus = "ok"
-	case active:
-		// toram was requested but medium is not yet/no longer in RAM
-		health.ToRAMStatus = "failed"
+	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: "toram boot parameter is set but the live medium is not mounted from RAM.",
+			Description: state.Message,
+		})
+	case "partial":
+		health.Issues = append(health.Issues, schema.RuntimeIssue{
+			Code:        "toram_copy_partial",
+			Severity:    "warning",
+			Description: state.Message,
 		})
-	default:
-		health.ToRAMStatus = "warning"
 	}
 }
 
@@ -211,13 +214,13 @@ func findUSBExportMount() string {
 
 	// 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,
+		"vfat":    true,
+		"exfat":   true,
+		"ext2":    true,
+		"ext3":    true,
+		"ext4":    true,
+		"ntfs":    true,
+		"ntfs3":   true,
 		"fuseblk": true,
 	}
 
@@ -244,11 +247,17 @@ func findUSBExportMount() string {
 		if readOnly {
 			continue
 		}
-		// Check USB transport via lsblk on the device
+		// Check USB transport via lsblk on the device (or its parent disk for partitions).
 		if !strings.HasPrefix(device, "/dev/") {
 			continue
 		}
-		if blockDeviceTransport(device) == "usb" {
+		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
 		}
 	}

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.
@@ -410,13 +410,13 @@ func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir strin
 	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
 		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
 		satJob{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
-			satJob{
-				name:       "03-dcgmproftester.log",
-				cmd:        profCmd,
-				env:        profEnv,
-				collectGPU: true,
-				gpuIndices: selected,
-			},
+		satJob{
+			name:       "03-dcgmproftester.log",
+			cmd:        profCmd,
+			env:        profEnv,
+			collectGPU: true,
+			gpuIndices: selected,
+		},
 		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
 	), logFunc)
 }
@@ -1382,8 +1382,6 @@ 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

@@ -426,6 +426,101 @@ 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) {
+	out, err := exec.Command("sensors", "-j").Output()
+	if err != nil || len(out) == 0 {
+		return 0, false
+	}
+	return parseFanDutyCyclePctSensorsJSON(out)
+}
+
+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

@@ -29,6 +29,27 @@ 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 TestParseDCMIPowerReading(t *testing.T) {
 	raw := `
 Instantaneous power reading:                   512 Watts

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

audit/internal/schema/hardware.go

@@ -15,17 +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"`
-	// ToRAMStatus: "ok" (ISO in RAM), "warning" (toram not active), "failed" (toram active but copy failed)
-	ToRAMStatus   string `json:"toram_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"`
+	USBExportPath string                 `json:"usb_export_path,omitempty"`
 	Issues        []RuntimeIssue         `json:"issues,omitempty"`
 	Tools         []RuntimeToolStatus    `json:"tools,omitempty"`
 	Services      []RuntimeServiceStatus `json:"services,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(),
 	}
@@ -482,13 +526,14 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			return
 		}
 
-			var body struct {
-				Duration           int      `json:"duration"`
-				StressMode         bool     `json:"stress_mode"`
-				GPUIndices         []int    `json:"gpu_indices"`
-				ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
-				StaggerGPUStart    bool     `json:"stagger_gpu_start"`
-				Loader             string   `json:"loader"`
+		var body struct {
+			Duration           int      `json:"duration"`
+			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"`
 			PlatformComponents []string `json:"platform_components"`
@@ -504,18 +549,153 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		if strings.TrimSpace(body.DisplayName) != "" {
 			name = body.DisplayName
 		}
-			params := taskParams{
-				Duration:           body.Duration,
-				StressMode:         body.StressMode,
-				GPUIndices:         body.GPUIndices,
-				ExcludeGPUIndices:  body.ExcludeGPUIndices,
-				StaggerGPUStart:    body.StaggerGPUStart,
-				Loader:             body.Loader,
+		params := taskParams{
+			Duration:           body.Duration,
+			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: resolve GPU list, then create one task per prefix
+			// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
+			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"))
+				var allTasks []*Task
+				for step := 1; step <= len(resolved); step++ {
+					subset := resolved[:step]
+					stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
+					t := &Task{
+						ID:        newJobID("bee-bench-nvidia"),
+						Name:      stepName,
+						Target:    target,
+						Priority:  defaultTaskPriority(target, taskParams{}),
+						Status:    TaskPending,
+						CreatedAt: now,
+						params: taskParams{
+							GPUIndices:       append([]int(nil), subset...),
+							SizeMB:           body.SizeMB,
+							BenchmarkProfile: body.Profile,
+							RunNCCL:          runNCCL && step == len(resolved),
+							ParallelGPUs:     true,
+							RampStep:         step,
+							RampTotal:        len(resolved),
+							RampRunID:        rampRunID,
+							DisplayName:      stepName,
+						},
+					}
+					allTasks = append(allTasks, t)
+				}
+				for _, t := range allTasks {
+					globalQueue.enqueue(t)
+				}
+				writeTaskRunResponse(w, allTasks)
+				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
@@ -528,56 +708,7 @@ 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) {
@@ -952,25 +1083,62 @@ 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(),
 	}
@@ -1085,7 +1253,7 @@ 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{
@@ -1377,4 +1545,3 @@ func (h *handler) rollbackPendingNetworkChange() error {
 	}
 	return nil
 }
-

audit/internal/webui/api_test.go

@@ -39,6 +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) {
@@ -61,7 +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)
@@ -75,8 +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)
@@ -84,6 +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) {
@@ -107,7 +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)
@@ -133,6 +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) {
@@ -175,6 +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/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/pages.go

@@ -330,6 +330,33 @@ func renderHardwareSummaryCard(opts HandlerOptions) string {
 
 	var b strings.Builder
 	b.WriteString(`<div class="card"><div class="card-head">Hardware Summary</div><div class="card-body">`)
+
+	// Server identity block above the component table.
+	{
+		var model, serial string
+		parts := []string{}
+		if hw.Board.Manufacturer != nil && strings.TrimSpace(*hw.Board.Manufacturer) != "" {
+			parts = append(parts, strings.TrimSpace(*hw.Board.Manufacturer))
+		}
+		if hw.Board.ProductName != nil && strings.TrimSpace(*hw.Board.ProductName) != "" {
+			parts = append(parts, strings.TrimSpace(*hw.Board.ProductName))
+		}
+		if len(parts) > 0 {
+			model = strings.Join(parts, " ")
+		}
+		serial = strings.TrimSpace(hw.Board.SerialNumber)
+		if model != "" || serial != "" {
+			b.WriteString(`<div style="margin-bottom:14px">`)
+			if model != "" {
+				fmt.Fprintf(&b, `<div style="font-size:16px;font-weight:700;margin-bottom:2px">%s</div>`, html.EscapeString(model))
+			}
+			if serial != "" {
+				fmt.Fprintf(&b, `<div style="font-size:12px;color:var(--muted)">S/N: %s</div>`, html.EscapeString(serial))
+			}
+			b.WriteString(`</div>`)
+		}
+	}
+
 	b.WriteString(`<table style="width:auto">`)
 	writeRow := func(label, value, badgeHTML string) {
 		b.WriteString(fmt.Sprintf(`<tr><td style="padding:6px 14px 6px 0;font-weight:700;white-space:nowrap">%s</td><td style="padding:6px 0;color:var(--muted);font-size:13px">%s</td><td style="padding:6px 0 6px 12px">%s</td></tr>`,
@@ -818,6 +845,13 @@ func buildRuntimeToRAMRow(health schema.RuntimeHealth) runtimeHealthRow {
 			Source: "live-boot / /proc/mounts",
 			Issue:  "",
 		}
+	case "partial":
+		return runtimeHealthRow{
+			Title:  "LiveCD in RAM",
+			Status: "WARNING",
+			Source: "live-boot / /proc/mounts / /dev/shm/bee-live",
+			Issue:  "Partial or staged RAM copy detected. System is not fully running from RAM; Copy to RAM can be retried.",
+		}
 	case "failed":
 		return runtimeHealthRow{
 			Title:  "LiveCD in RAM",
@@ -1279,9 +1313,6 @@ func renderValidate(opts HandlerOptions) string {
 	<div class="card" style="margin-bottom:16px">
 	  <div class="card-head">Validate Profile</div>
 	  <div class="card-body validate-profile-body">
-	    <div class="validate-profile-col">
-	      <div class="form-row" style="margin:0"><label>Cycles</label><input type="number" id="sat-cycles" value="1" min="1" max="100" style="width:100%"></div>
-	    </div>
 	    <div class="validate-profile-col">
 	      <div class="form-row" style="margin:12px 0 0"><label>Mode</label></div>
 	      <label class="cb-row"><input type="radio" name="sat-mode" id="sat-mode-validate" value="validate" checked onchange="satModeChanged()"><span>Validate — quick non-destructive check</span></label>
@@ -1331,12 +1362,6 @@ func renderValidate(opts HandlerOptions) string {
       <p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs...</p>
     </div>
     <p id="sat-gpu-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 0">Select at least one NVIDIA GPU to enable NVIDIA validate tasks.</p>
-    <div style="margin-top:10px;padding-top:10px;border-top:1px solid var(--border)">
-      <label class="sat-gpu-row" title="When checked, multi-GPU tests (PSU Pulse, NCCL, NVBandwidth) run on ALL GPUs in the system regardless of the selection above.">
-        <input type="checkbox" id="sat-multi-gpu-all" checked onchange="satUpdateGPUSelectionNote()">
-        <span><strong>Multi-GPU tests</strong> — use all GPUs <span style="font-size:11px;color:var(--muted)">(PSU Pulse, NCCL, NVBandwidth)</span></span>
-      </label>
-    </div>
   </div>
 </div>
 
@@ -1455,10 +1480,6 @@ function satSelectedGPUIndices() {
     .filter(function(v) { return !Number.isNaN(v); })
     .sort(function(a, b) { return a - b; });
 }
-function satMultiGPUAll() {
-  const cb = document.getElementById('sat-multi-gpu-all');
-  return cb ? cb.checked : true;
-}
 function satUpdateGPUSelectionNote() {
   const note = document.getElementById('sat-gpu-selection-note');
   if (!note) return;
@@ -1467,8 +1488,7 @@ function satUpdateGPUSelectionNote() {
     note.textContent = 'Select at least one NVIDIA GPU to enable NVIDIA validate tasks.';
     return;
   }
-  const multiAll = satMultiGPUAll();
-  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '. Multi-GPU tests: ' + (multiAll ? 'all GPUs in system' : 'selected GPUs only') + '.';
+  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '. Multi-GPU tests will use all selected GPUs.';
 }
 function satRenderGPUList(gpus) {
   const root = document.getElementById('sat-gpu-list');
@@ -1582,15 +1602,8 @@ const nvidiaPerGPUTargets = ['nvidia', 'nvidia-targeted-stress', 'nvidia-targete
 // pulse_test and fabric tests run on all selected GPUs simultaneously
 const nvidiaAllGPUTargets = ['nvidia-pulse', 'nvidia-interconnect', 'nvidia-bandwidth'];
 function satAllGPUIndicesForMulti() {
-  // If "Multi-GPU tests — all GPUs" is checked, return all detected GPUs.
-  // Otherwise fall back to the per-GPU selection.
-  if (satMultiGPUAll()) {
-    return loadSatNvidiaGPUs().then(function(gpus) {
-      return gpus.map(function(g) { return Number(g.index); });
-    });
-  }
-  const sel = satSelectedGPUIndices();
-  return Promise.resolve(sel);
+  // Multi-GPU tests always use the current GPU selection.
+  return Promise.resolve(satSelectedGPUIndices());
 }
 function expandSATTarget(target) {
   if (nvidiaAllGPUTargets.indexOf(target) >= 0) {
@@ -1680,7 +1693,7 @@ function runAMDValidateSet() {
   return runNext(0);
 }
 function runAllSAT() {
-  const cycles = Math.max(1, parseInt(document.getElementById('sat-cycles').value)||1);
+  const cycles = 1;
   const status = document.getElementById('sat-all-status');
   status.textContent = 'Enqueuing...';
   const stressOnlyTargets = ['nvidia-targeted-stress', 'nvidia-targeted-power', 'nvidia-pulse', 'nvidia-interconnect', 'nvidia-bandwidth'];
@@ -1922,23 +1935,10 @@ func renderSATCard(id, label, runAction, headerActions, body string) string {
 
 // ── Benchmark ─────────────────────────────────────────────────────────────────
 
-type benchmarkHistoryColumn struct {
-	key      string
-	label    string
-	name     string
-	index    int
-	parallel bool
-}
-
-type benchmarkHistoryCell struct {
-	score   float64
-	present bool
-}
-
 type benchmarkHistoryRun struct {
 	generatedAt time.Time
 	displayTime string
-	cells       map[string]benchmarkHistoryCell
+	gpuScores   map[int]float64 // GPU index → composite score
 }
 
 func renderBenchmark(opts HandlerOptions) string {
@@ -1946,7 +1946,7 @@ func renderBenchmark(opts HandlerOptions) string {
 
 <div class="grid2">
   <div class="card">
-    <div class="card-head">NVIDIA Benchmark</div>
+    <div class="card-head">Benchmark Setup</div>
     <div class="card-body">
       <div class="form-row">
         <label>Profile</label>
@@ -1967,29 +1967,37 @@ func renderBenchmark(opts HandlerOptions) string {
         </div>
       </div>
       <label class="benchmark-cb-row">
-        <input type="checkbox" id="benchmark-parallel-gpus">
-        <span>Run all selected GPUs simultaneously (parallel mode)</span>
+        <input type="radio" name="benchmark-mode" value="sequential" onchange="benchmarkUpdateSelectionNote()">
+        <span>Sequential — one GPU at a time</span>
       </label>
-      <label class="benchmark-cb-row">
-        <input type="checkbox" id="benchmark-run-nccl" checked>
-        <span>Run multi-GPU interconnect step (NCCL) only on the selected GPUs</span>
+      <label class="benchmark-cb-row" id="benchmark-parallel-label">
+        <input type="radio" name="benchmark-mode" value="parallel" onchange="benchmarkUpdateSelectionNote()">
+        <span>Parallel — all selected GPUs simultaneously</span>
+      </label>
+      <label class="benchmark-cb-row" id="benchmark-ramp-label">
+        <input type="radio" name="benchmark-mode" value="ramp-up" checked onchange="benchmarkUpdateSelectionNote()">
+        <span>Ramp-up — 1 GPU → 2 → … → all selected (separate tasks)</span>
       </label>
       <p id="benchmark-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 14px">Select one GPU for single-card benchmarking or several GPUs for a constrained multi-GPU run.</p>
-      <button id="benchmark-run-btn" class="btn btn-primary" onclick="runNvidiaBenchmark()" disabled>&#9654; Run Benchmark</button>
+      <div style="display:flex;gap:8px;flex-wrap:wrap;align-items:center">
+        <button id="benchmark-run-performance-btn" class="btn btn-primary" onclick="runNvidiaBenchmark('performance')" disabled>&#9654; Run Performance Benchmark</button>
+        <button id="benchmark-run-power-fit-btn" class="btn btn-secondary" onclick="runNvidiaBenchmark('power-fit')" disabled>&#9654; Run Power / Thermal Fit</button>
+      </div>
+      <span id="benchmark-run-nccl" hidden>nccl-auto</span>
       <span id="benchmark-run-status" style="margin-left:10px;font-size:12px;color:var(--muted)"></span>
     </div>
   </div>
 
   <div class="card">
-    <div class="card-head">Method</div>
+    <div class="card-head">Method Split</div>
     <div class="card-body">
-      <p style="font-size:13px;color:var(--muted);margin-bottom:10px">Each benchmark run performs warmup, sustained compute, telemetry capture, cooldown, and optional NCCL interconnect checks.</p>
+      <p style="font-size:13px;color:var(--muted);margin-bottom:10px">The benchmark page now exposes two fundamentally different test families so compute score and server power-fit are not mixed into one number.</p>
       <table>
-        <tr><th>Profile</th><th>Purpose</th></tr>
-        <tr><td>Standard</td><td>Fast, repeatable performance check for server-to-server comparison.</td></tr>
-        <tr><td>Stability</td><td>Longer run for thermal drift, power caps, and clock instability.</td></tr>
-        <tr><td>Overnight</td><td>Extended verification of long-run stability and late throttling.</td></tr>
+        <tr><th>Run Type</th><th>Engine</th><th>Question</th></tr>
+        <tr><td>Performance Benchmark</td><td><code>bee-gpu-burn</code></td><td>How much isolated compute performance does the GPU realize in this server?</td></tr>
+        <tr><td>Power / Thermal Fit</td><td><code>dcgmi targeted_power</code></td><td>How much power per GPU can this server sustain as GPU count ramps up?</td></tr>
       </table>
+      <p style="font-size:12px;color:var(--muted);margin-top:10px">Use ramp-up mode for capacity work: it creates 1 GPU → 2 GPU → … → all selected steps so analysis software can derive server total score and watts-per-GPU curves.</p>
     </div>
   </div>
 </div>
@@ -2025,22 +2033,31 @@ function benchmarkSelectedGPUIndices() {
     .sort(function(a, b) { return a - b; });
 }
 
+function benchmarkMode() {
+  const el = document.querySelector('input[name="benchmark-mode"]:checked');
+  return el ? el.value : 'sequential';
+}
+
 function benchmarkUpdateSelectionNote() {
   const selected = benchmarkSelectedGPUIndices();
-  const btn = document.getElementById('benchmark-run-btn');
+  const perfBtn = document.getElementById('benchmark-run-performance-btn');
+  const fitBtn = document.getElementById('benchmark-run-power-fit-btn');
   const note = document.getElementById('benchmark-selection-note');
-  const nccl = document.getElementById('benchmark-run-nccl');
   if (!selected.length) {
-    btn.disabled = true;
+    perfBtn.disabled = true;
+    fitBtn.disabled = true;
     note.textContent = 'Select at least one NVIDIA GPU to run the benchmark.';
     return;
   }
-  btn.disabled = false;
-  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '.';
-  if (nccl && nccl.checked && selected.length < 2) {
-    note.textContent += ' NCCL will be skipped because fewer than 2 GPUs are selected.';
-  } else if (nccl && nccl.checked) {
-    note.textContent += ' NCCL interconnect will use only these GPUs.';
+  perfBtn.disabled = false;
+  fitBtn.disabled = false;
+  const mode = benchmarkMode();
+  if (mode === 'ramp-up') {
+    note.textContent = 'Ramp-up: ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). Performance uses compute benchmark; Power / Thermal Fit uses targeted_power per step.';
+  } else if (mode === 'parallel') {
+    note.textContent = 'Parallel: all ' + selected.length + ' GPU(s) simultaneously. Only the performance benchmark supports this mode.';
+  } else {
+    note.textContent = 'Sequential: each selected GPU benchmarked separately.';
   }
 }
 
@@ -2058,6 +2075,33 @@ function benchmarkRenderGPUList(gpus) {
       + '<span><strong>GPU ' + gpu.index + '</strong> — ' + gpu.name + mem + '</span>'
       + '</label>';
   }).join('');
+  benchmarkApplyMultiGPUState(gpus.length);
+  benchmarkUpdateSelectionNote();
+}
+
+// Disable radio options that require multiple GPUs when only one is present.
+function benchmarkApplyMultiGPUState(gpuCount) {
+  var multiValues = ['parallel', 'ramp-up'];
+  var radios = document.querySelectorAll('input[name="benchmark-mode"]');
+  radios.forEach(function(el) {
+    var isMulti = multiValues.indexOf(el.value) >= 0;
+    if (gpuCount < 2 && isMulti) {
+      el.disabled = true;
+      if (el.checked) {
+        // fall back to sequential
+        var seq = document.querySelector('input[name="benchmark-mode"][value="sequential"]');
+        if (seq) seq.checked = true;
+      }
+      var label = el.closest('label');
+      if (label) label.style.opacity = '0.4';
+    } else {
+      el.disabled = false;
+      // restore default: ramp-up checked when ≥2 GPUs
+      if (gpuCount >= 2 && el.value === 'ramp-up') el.checked = true;
+      var label = el.closest('label');
+      if (label) label.style.opacity = '';
+    }
+  });
   benchmarkUpdateSelectionNote();
 }
 
@@ -2087,7 +2131,7 @@ function benchmarkSelectNone() {
   benchmarkUpdateSelectionNote();
 }
 
-function runNvidiaBenchmark() {
+function runNvidiaBenchmark(kind) {
   const selected = benchmarkSelectedGPUIndices();
   const status = document.getElementById('benchmark-run-status');
   if (!selected.length) {
@@ -2095,20 +2139,28 @@ function runNvidiaBenchmark() {
     return;
   }
   if (benchmarkES) { benchmarkES.close(); benchmarkES = null; }
-  const parallelGPUs = !!document.getElementById('benchmark-parallel-gpus').checked;
+  const mode = benchmarkMode();
+  const rampUp = mode === 'ramp-up' && selected.length > 1;
+  const parallelGPUs = mode === 'parallel' && kind === 'performance';
+  if (kind === 'power-fit' && mode === 'parallel') {
+    status.textContent = 'Power / Thermal Fit supports sequential or ramp-up only.';
+    return;
+  }
   const body = {
     profile: document.getElementById('benchmark-profile').value || 'standard',
     gpu_indices: selected,
-    run_nccl: !!document.getElementById('benchmark-run-nccl').checked,
+    run_nccl: kind === 'performance' && selected.length > 1,
     parallel_gpus: parallelGPUs,
-    display_name: 'NVIDIA Benchmark'
+    ramp_up: rampUp,
+    display_name: kind === 'power-fit' ? 'NVIDIA Power / Thermal Fit' : 'NVIDIA Performance Benchmark'
   };
   document.getElementById('benchmark-output').style.display = 'block';
-  document.getElementById('benchmark-title').textContent = '— ' + body.profile + ' [' + selected.join(', ') + ']';
+  document.getElementById('benchmark-title').textContent = '— ' + body.display_name + ' · ' + body.profile + ' [' + selected.join(', ') + ']';
   const term = document.getElementById('benchmark-terminal');
-  term.textContent = 'Enqueuing benchmark for GPUs ' + selected.join(', ') + '...\n';
+  term.textContent = 'Enqueuing ' + body.display_name + ' for GPUs ' + selected.join(', ') + '...\n';
   status.textContent = 'Queueing...';
-  fetch('/api/benchmark/nvidia/run', {
+  const endpoint = kind === 'power-fit' ? '/api/bee-bench/nvidia/power/run' : '/api/bee-bench/nvidia/perf/run';
+  fetch(endpoint, {
     method: 'POST',
     headers: {'Content-Type':'application/json'},
     body: JSON.stringify(body)
@@ -2155,23 +2207,22 @@ function runNvidiaBenchmark() {
   });
 }
 
-document.getElementById('benchmark-run-nccl').addEventListener('change', benchmarkUpdateSelectionNote);
 benchmarkLoadGPUs();
 </script>`
 }
 
 func renderBenchmarkResultsCard(exportDir string) string {
-	columns, runs := loadBenchmarkHistory(exportDir)
+	maxIdx, runs := loadBenchmarkHistory(exportDir)
 	return renderBenchmarkResultsCardFromRuns(
-		"Benchmark Results",
+		"Perf Results",
 		"Composite score by saved benchmark run and GPU.",
 		"No saved benchmark runs yet.",
-		columns,
+		maxIdx,
 		runs,
 	)
 }
 
-func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, columns []benchmarkHistoryColumn, runs []benchmarkHistoryRun) string {
+func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, maxGPUIndex int, runs []benchmarkHistoryRun) string {
 	if len(runs) == 0 {
 		return `<div class="card"><div class="card-head">` + html.EscapeString(title) + `</div><div class="card-body"><p style="color:var(--muted);font-size:13px">` + html.EscapeString(emptyMessage) + `</p></div></div>`
 	}
@@ -2181,22 +2232,22 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 		b.WriteString(`<p style="color:var(--muted);font-size:13px;margin-bottom:12px">` + html.EscapeString(description) + `</p>`)
 	}
 	b.WriteString(`<div style="overflow-x:auto">`)
-	b.WriteString(`<table><thead><tr><th>Test</th><th>Time</th>`)
-	for _, col := range columns {
-		b.WriteString(`<th>` + html.EscapeString(col.label) + `</th>`)
+	b.WriteString(`<table><thead><tr><th>Run</th><th>Time</th>`)
+	for i := 0; i <= maxGPUIndex; i++ {
+		b.WriteString(`<th>GPU ` + strconv.Itoa(i) + `</th>`)
 	}
 	b.WriteString(`</tr></thead><tbody>`)
 	for i, run := range runs {
 		b.WriteString(`<tr>`)
 		b.WriteString(`<td>#` + strconv.Itoa(i+1) + `</td>`)
 		b.WriteString(`<td>` + html.EscapeString(run.displayTime) + `</td>`)
-		for _, col := range columns {
-			cell, ok := run.cells[col.key]
-			if !ok || !cell.present {
+		for idx := 0; idx <= maxGPUIndex; idx++ {
+			score, ok := run.gpuScores[idx]
+			if !ok {
 				b.WriteString(`<td style="color:var(--muted)">-</td>`)
 				continue
 			}
-			b.WriteString(`<td>` + fmt.Sprintf("%.2f", cell.score) + `</td>`)
+			b.WriteString(`<td>` + fmt.Sprintf("%.2f", score) + `</td>`)
 		}
 		b.WriteString(`</tr>`)
 	}
@@ -2204,22 +2255,22 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 	return b.String()
 }
 
-func loadBenchmarkHistory(exportDir string) ([]benchmarkHistoryColumn, []benchmarkHistoryRun) {
-	baseDir := app.DefaultBenchmarkBaseDir
+func loadBenchmarkHistory(exportDir string) (int, []benchmarkHistoryRun) {
+	baseDir := app.DefaultBeeBenchPerfDir
 	if strings.TrimSpace(exportDir) != "" {
-		baseDir = filepath.Join(exportDir, "bee-benchmark")
+		baseDir = filepath.Join(exportDir, "bee-bench", "perf")
 	}
-	paths, err := filepath.Glob(filepath.Join(baseDir, "gpu-benchmark-*", "result.json"))
+	paths, err := filepath.Glob(filepath.Join(baseDir, "perf-*", "result.json"))
 	if err != nil || len(paths) == 0 {
-		return nil, nil
+		return -1, nil
 	}
 	sort.Strings(paths)
 	return loadBenchmarkHistoryFromPaths(paths)
 }
 
-func loadBenchmarkHistoryFromPaths(paths []string) ([]benchmarkHistoryColumn, []benchmarkHistoryRun) {
-	columnByKey := make(map[string]benchmarkHistoryColumn)
+func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun) {
 	runs := make([]benchmarkHistoryRun, 0, len(paths))
+	maxGPUIndex := -1
 	for _, path := range paths {
 		raw, err := os.ReadFile(path)
 		if err != nil {
@@ -2232,101 +2283,20 @@ func loadBenchmarkHistoryFromPaths(paths []string) ([]benchmarkHistoryColumn, []
 		run := benchmarkHistoryRun{
 			generatedAt: result.GeneratedAt,
 			displayTime: result.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
-			cells:       make(map[string]benchmarkHistoryCell),
+			gpuScores:   make(map[int]float64),
 		}
-
-		if result.ParallelGPUs {
-			// All GPUs ran simultaneously — one column per server, score = avg composite.
-			gpuModelCount := make(map[string]int)
-			for _, gpu := range result.GPUs {
-				gpuModelCount[strings.TrimSpace(gpu.Name)]++
-			}
-			scoreSum := make(map[string]float64)
-			scoreCnt := make(map[string]int)
-			for _, gpu := range result.GPUs {
-				key := "parallel|" + strings.TrimSpace(result.ServerModel) + "|" + strings.TrimSpace(gpu.Name)
-				scoreSum[key] += gpu.Scores.CompositeScore
-				scoreCnt[key]++
-				count := gpuModelCount[strings.TrimSpace(gpu.Name)]
-				columnByKey[key] = benchmarkHistoryColumn{
-					key:      key,
-					label:    benchmarkHistoryParallelLabel(result.ServerModel, gpu.Name, count),
-					name:     strings.TrimSpace(gpu.Name),
-					index:    -1,
-					parallel: true,
-				}
-			}
-			for key, sum := range scoreSum {
-				run.cells[key] = benchmarkHistoryCell{score: sum / float64(scoreCnt[key]), present: true}
-			}
-		} else {
-			// Each GPU ran independently — one column per GPU index.
-			for _, gpu := range result.GPUs {
-				key := "gpu|" + strings.TrimSpace(result.ServerModel) + "|" + strings.TrimSpace(gpu.Name) + "|" + strconv.Itoa(gpu.Index)
-				columnByKey[key] = benchmarkHistoryColumn{
-					key:      key,
-					label:    benchmarkHistoryPerGPULabel(gpu.Name, gpu.Index),
-					name:     strings.TrimSpace(gpu.Name),
-					index:    gpu.Index,
-					parallel: false,
-				}
-				run.cells[key] = benchmarkHistoryCell{score: gpu.Scores.CompositeScore, present: true}
+		for _, gpu := range result.GPUs {
+			run.gpuScores[gpu.Index] = gpu.Scores.CompositeScore
+			if gpu.Index > maxGPUIndex {
+				maxGPUIndex = gpu.Index
 			}
 		}
 		runs = append(runs, run)
 	}
-
-	columns := make([]benchmarkHistoryColumn, 0, len(columnByKey))
-	for _, col := range columnByKey {
-		columns = append(columns, col)
-	}
-	// Sequential GPU columns first (sorted by GPU index), then parallel server columns.
-	sort.Slice(columns, func(i, j int) bool {
-		if columns[i].parallel != columns[j].parallel {
-			return !columns[i].parallel // sequential first
-		}
-		if columns[i].parallel {
-			li := strings.ToLower(columns[i].label)
-			lj := strings.ToLower(columns[j].label)
-			if li != lj {
-				return li < lj
-			}
-			return columns[i].key < columns[j].key
-		}
-		// Sequential: sort by GPU index, then name.
-		if columns[i].index != columns[j].index {
-			return columns[i].index < columns[j].index
-		}
-		return strings.ToLower(columns[i].name) < strings.ToLower(columns[j].name)
-	})
 	sort.Slice(runs, func(i, j int) bool {
 		return runs[i].generatedAt.After(runs[j].generatedAt)
 	})
-	return columns, runs
-}
-
-// benchmarkHistoryPerGPULabel formats a label for a single-GPU column: "GPU #N — ModelName".
-func benchmarkHistoryPerGPULabel(gpuName string, index int) string {
-	gpuName = strings.TrimSpace(gpuName)
-	if gpuName == "" {
-		gpuName = "Unknown GPU"
-	}
-	return fmt.Sprintf("GPU #%d — %s", index, gpuName)
-}
-
-// benchmarkHistoryParallelLabel formats a label for an all-GPU parallel column:
-// "ServerModel — N× ModelName (All GPUs)" or "N× ModelName (All GPUs)" if no server.
-func benchmarkHistoryParallelLabel(serverModel, gpuName string, count int) string {
-	serverModel = strings.TrimSpace(serverModel)
-	gpuName = strings.TrimSpace(gpuName)
-	if gpuName == "" {
-		gpuName = "Unknown GPU"
-	}
-	gpuPart := fmt.Sprintf("%d× %s (All GPUs)", count, gpuName)
-	if serverModel == "" {
-		return gpuPart
-	}
-	return fmt.Sprintf("%s — %s", serverModel, gpuPart)
+	return maxGPUIndex, runs
 }
 
 // ── Burn ──────────────────────────────────────────────────────────────────────
@@ -2371,10 +2341,20 @@ func renderBurn() string {
 	      <p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs...</p>
 	    </div>
 	    <p id="burn-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 0">Select at least one NVIDIA GPU to enable NVIDIA burn recipes.</p>
-	    <label class="cb-row" style="margin-top:10px">
-	      <input type="checkbox" id="burn-stagger-nvidia">
-	      <span>Ramp selected NVIDIA GPUs one by one before the full-load hold. Smoke: +2 min per GPU, then 5 min with all selected GPUs under load. Acceptance: +10 min per GPU, then at least 1 hour with all selected GPUs under load. Overnight: +1 hour per GPU, then at least 1 hour with all selected GPUs under load, capped at 10 hours total.</span>
-	    </label>
+	    <div style="display:flex;flex-direction:column;gap:4px;margin-top:10px">
+	      <label class="cb-row">
+	        <input type="radio" name="burn-nvidia-mode" value="sequential" checked>
+	        <span>Sequential — selected GPUs one at a time</span>
+	      </label>
+	      <label class="cb-row" id="burn-parallel-label">
+	        <input type="radio" name="burn-nvidia-mode" value="parallel">
+	        <span>Parallel — all selected GPUs simultaneously</span>
+	      </label>
+	      <label class="cb-row" id="burn-ramp-label">
+	        <input type="radio" name="burn-nvidia-mode" value="ramp-up">
+	        <span>Ramp-up — add one GPU at a time</span>
+	      </label>
+	    </div>
 	  </div>
 	</div>
 
@@ -2450,9 +2430,30 @@ function burnSelectedGPUIndices() {
     .sort(function(a, b) { return a - b; });
 }
 
-function burnUseNvidiaRampUp() {
-  const el = document.getElementById('burn-stagger-nvidia');
-  return !!(el && el.checked);
+function burnNvidiaMode() {
+  const el = document.querySelector('input[name="burn-nvidia-mode"]:checked');
+  return el ? el.value : 'sequential';
+}
+
+function burnApplyMultiGPUState(gpuCount) {
+  var multiValues = ['parallel', 'ramp-up'];
+  var radios = document.querySelectorAll('input[name="burn-nvidia-mode"]');
+  radios.forEach(function(el) {
+    var isMulti = multiValues.indexOf(el.value) >= 0;
+    if (gpuCount < 2 && isMulti) {
+      el.disabled = true;
+      if (el.checked) {
+        var seq = document.querySelector('input[name="burn-nvidia-mode"][value="sequential"]');
+        if (seq) seq.checked = true;
+      }
+      var label = el.closest('label');
+      if (label) label.style.opacity = '0.4';
+    } else {
+      el.disabled = false;
+      var label = el.closest('label');
+      if (label) label.style.opacity = '';
+    }
+  });
 }
 
 function burnUpdateSelectionNote() {
@@ -2479,6 +2480,7 @@ function burnRenderGPUList(gpus) {
       + '<span><strong>GPU ' + gpu.index + '</strong> — ' + gpu.name + mem + '</span>'
       + '</label>';
   }).join('');
+  burnApplyMultiGPUState(gpus.length);
   burnUpdateSelectionNote();
 }
 
@@ -2514,8 +2516,11 @@ function enqueueBurnTask(target, label, extra, useSelectedNvidia) {
       return Promise.reject(new Error('Select at least one NVIDIA GPU.'));
     }
     body.gpu_indices = selected;
-    if (burnUseNvidiaRampUp() && selected.length > 1) {
+    const bMode = burnNvidiaMode();
+    if (bMode === 'ramp-up' && selected.length > 1) {
       body.stagger_gpu_start = true;
+    } else if (bMode === 'parallel' && selected.length > 1) {
+      body.parallel_gpus = true;
     }
   }
   return fetch('/api/sat/' + target + '/run', {
@@ -3258,12 +3263,19 @@ fetch('/api/system/ram-status').then(r=>r.json()).then(d=>{
   else if (kind === 'disk') label = 'disk (' + source + ')';
   else label = source;
   boot.textContent = 'Current boot source: ' + label + '.';
-  if (d.in_ram) {
-    txt.textContent = '✓ Running from RAM — installation media can be safely disconnected.';
+  txt.textContent = d.message || 'Checking...';
+  if (d.status === 'ok' || d.in_ram) {
     txt.style.color = 'var(--ok, green)';
+  } else if (d.status === 'failed') {
+    txt.style.color = 'var(--err, #b91c1c)';
   } else {
-    txt.textContent = 'Live media is mounted from installation device. Copy to RAM to allow media removal.';
+    txt.style.color = 'var(--muted)';
+  }
+  if (d.can_start_task) {
     btn.style.display = '';
+    btn.disabled = false;
+  } else {
+    btn.style.display = 'none';
   }
 });
 function installToRAM() {

audit/internal/webui/server.go

@@ -261,7 +261,8 @@ 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"))
 
 	// Tasks
 	mux.HandleFunc("GET /api/tasks", h.handleAPITasksList)

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`,
@@ -1113,8 +1127,8 @@ func TestDashboardRendersRuntimeHealthTable(t *testing.T) {
 		`>Storage<`,
 		`>GPU<`,
 		`>PSU<`,
-		`badge-warn`,   // cpu Warning badge
-		`badge-err`,    // storage Critical badge
+		`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

@@ -32,7 +32,8 @@ 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)",
@@ -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"`
@@ -625,7 +629,7 @@ 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
@@ -637,6 +641,22 @@ 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 {

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) {

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.
 

iso/builder/VERSIONS

@@ -6,7 +6,7 @@ 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

iso/builder/bee-gpu-stress.c

@@ -33,7 +33,6 @@ 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)
 
@@ -643,6 +642,20 @@ 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 +702,8 @@ static const struct profile_desc k_profiles[] = {
 #endif
 };
 
+#define PROFILE_COUNT ((int)(sizeof(k_profiles) / sizeof(k_profiles[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 +774,12 @@ 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 +792,16 @@ 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;
@@ -943,8 +970,9 @@ static int prepare_profile(struct cublaslt_api *cublas,
         return 0;
     }
 
+    cudaDataType_t scale_type = matmul_scale_type(desc);
     if (!check_cublas("cublasLtMatmulDescCreate",
-                      cublas->cublasLtMatmulDescCreate(&out->op_desc, desc->compute_type, CUDA_R_32F))) {
+                      cublas->cublasLtMatmulDescCreate(&out->op_desc, desc->compute_type, scale_type))) {
         destroy_profile(cublas, cuda, out);
         return 0;
     }
@@ -1093,17 +1121,30 @@ static int prepare_profile(struct cublaslt_api *cublas,
 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 = α
+    const void *beta_ptr = β
+    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 +1162,10 @@ 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};
@@ -1133,7 +1175,7 @@ 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;
@@ -1158,8 +1200,10 @@ 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 (k_profiles[i].enabled && cc >= k_profiles[i].min_cc &&
+            (precision_filter == NULL || strcmp(k_profiles[i].block_label, precision_filter) == 0)) {
             planned++;
         }
     }
@@ -1170,18 +1214,31 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Count all profiles active on this GPU regardless of filter.
+     * Used as the budget divisor so matrix sizes stay consistent whether
+     * running all precisions together or a single-precision phase. */
+    int planned_total = 0;
+    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) {
+            planned_total++;
+        }
+    }
+    if (planned_total < planned) {
+        planned_total = planned;
+    }
+
     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)planned_total * MIN_PROFILE_BUDGET_BYTES) {
+        requested_budget = (size_t)planned_total * 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)planned_total * MIN_PROFILE_BUDGET_BYTES) {
+        total_budget = (size_t)planned_total * 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, planned_total, total_budget, 1);
     }
     if (stream_count > 1) {
         int created = 0;
@@ -1194,7 +1251,7 @@ 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)planned_total * (size_t)stream_count);
     if (per_profile_budget < MIN_PROFILE_BUDGET_BYTES) {
         per_profile_budget = MIN_PROFILE_BUDGET_BYTES;
     }
@@ -1218,6 +1275,13 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                           desc->min_cc);
             continue;
         }
+        if (precision_filter != NULL && strcmp(desc->block_label, precision_filter) != 0) {
+            append_detail(report->details,
+                          sizeof(report->details),
+                          "%s=SKIPPED precision_filter\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 +1399,29 @@ 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 +1429,16 @@ 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 +1498,94 @@ 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.sh

@@ -873,9 +873,37 @@ 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 [ "$GPU_STRESS_NEED_BUILD" = "1" ]; then
@@ -889,6 +917,12 @@ if [ "$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}) ==="

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

@@ -6,10 +6,13 @@ 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] [--stagger-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
 }
 
@@ -30,6 +33,9 @@ while [ "$#" -gt 0 ]; do
         --size-mb|-m) [ "$#" -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
@@ -88,8 +94,14 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
     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}" \
-        "${WORKER}" --device 0 --seconds "${gpu_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