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>
Add slot-aware ramp sequence to bee-bench power
2026-04-14 20:41:17 +03:00 · 2026-04-14 17:47:40 +03:00 · 2026-04-14 17:33:13 +03:00 · 2026-04-14 16:18:33 +03:00 · 2026-04-14 14:12:06 +03:00
24 changed files with 1827 additions and 449 deletions
--- a/audit/internal/app/app.go
+++ b/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
--- a/audit/internal/app/app_test.go
+++ b/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)
--- a/audit/internal/platform/benchmark.go
+++ b/audit/internal/platform/benchmark.go
--- a/audit/internal/platform/benchmark_report.go
+++ b/audit/internal/platform/benchmark_report.go
@@ -48,7 +48,7 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		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.RampStep > 0 && result.RampTotal > 0 {
 		fmt.Fprintf(&b, "**Ramp-up step:** %d of %d  \n", result.RampStep, result.RampTotal)
@@ -83,15 +83,15 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {

 	// ── Methodology ───────────────────────────────────────────────────────────
 	b.WriteString("## Methodology\n\n")
-	fmt.Fprintf(&b, "- Profile `%s` uses standardized baseline -> warmup -> steady-state -> interconnect -> cooldown phases.\n", result.BenchmarkProfile)
+	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 (fp8, fp16, fp32, fp64, fp4) runs alone for a dedicated window. ")
+	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 · fp8 ×0.25 · fp4 ×0.125.\n")
+	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")
@@ -170,6 +170,16 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		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)
 		}
@@ -188,7 +198,7 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		// Per-precision stability phases.
 		if len(gpu.PrecisionSteady) > 0 {
 			b.WriteString("**Per-precision stability:**\n\n")
-			b.WriteString("| Precision | Clock CV | Power CV | Clock Drift | ECC corr | ECC uncorr |\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 := "—"
@@ -196,8 +206,12 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 					eccCorr = fmt.Sprintf("%d", p.ECC.Corrected)
 					eccUncorr = fmt.Sprintf("%d", p.ECC.Uncorrected)
 				}
-				fmt.Fprintf(&b, "| %s | %.1f%% | %.1f%% | %.1f%% | %s | %s |\n",
-					p.Precision, p.Steady.ClockCVPct, p.Steady.PowerCVPct, p.Steady.ClockDriftPct,
+				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")
@@ -364,6 +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))
--- a/audit/internal/platform/benchmark_test.go
+++ b/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) {
--- a/audit/internal/platform/benchmark_types.go
+++ b/audit/internal/platform/benchmark_types.go
@@ -104,6 +104,7 @@ type BenchmarkGPUResult struct {
 	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
@@ -111,6 +112,8 @@ type BenchmarkGPUResult struct {
 	// 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"`
@@ -119,6 +122,7 @@ type BenchmarkGPUResult struct {
 	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).
@@ -179,7 +183,7 @@ type BenchmarkPrecisionResult struct {
 	Iterations    uint64  `json:"iterations,omitempty"`
 	TeraOpsPerSec float64 `json:"teraops_per_sec,omitempty"`
 	// Weight is the fp32-equivalence factor for this precision category.
-	// fp32 = 1.0 (baseline), fp64 = 2.0, fp16 = 0.5, fp8 = 0.25, fp4 = 0.125.
+	// 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"`
@@ -225,13 +229,15 @@ type BenchmarkServerPower struct {
 // 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"`
+	ECC   BenchmarkECCCounters `json:"ecc,omitempty"`
+	Notes string               `json:"notes,omitempty"`
 }

 type BenchmarkInterconnectResult struct {
@@ -245,3 +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"`
+}
--- a/audit/internal/platform/gpu_metrics.go
+++ b/audit/internal/platform/gpu_metrics.go
@@ -14,6 +14,8 @@ import (
 // GPUMetricRow is one telemetry sample from nvidia-smi during a stress test.
 type GPUMetricRow struct {
 	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"`
@@ -509,11 +511,22 @@ func buildGPUMetricStageSpans(rows []GPUMetricRow) []gpuMetricStageSpan {
 		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: row.ElapsedSec, End: row.ElapsedSec})
+			spans = append(spans, gpuMetricStageSpan{Name: name, Start: start, End: end})
 			continue
 		}
-		spans[len(spans)-1].End = row.ElapsedSec
+		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 {
--- a/audit/internal/platform/install_to_ram.go
+++ b/audit/internal/platform/install_to_ram.go
@@ -11,20 +11,10 @@ import (
 	"strings"
 )

+const installToRAMDir = "/dev/shm/bee-live"
+
 func (s *System) IsLiveMediaInRAM() bool {
-	fsType := mountFSType("/run/live/medium")
-	if fsType == "" {
-		// No medium mount at all — fall back to toram kernel parameter.
-		return toramActive()
-	}
-	if strings.EqualFold(fsType, "tmpfs") {
-		return true
-	}
-	// When RunInstallToRAM copies squashfs to /dev/shm/bee-live but the bind
-	// mount of /run/live/medium fails (common for CD-ROM boots), the medium
-	// fstype still shows the CD-ROM type. Check whether the RAM copy exists.
-	files, _ := filepath.Glob("/dev/shm/bee-live/*.squashfs")
-	return len(files) > 0
+	return s.LiveMediaRAMState().InRAM
 }

 func (s *System) LiveBootSource() LiveBootSource {
@@ -56,14 +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
 	}
@@ -88,10 +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 {
--- a/audit/internal/platform/install_to_ram_test.go
+++ b/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)
+		}
+	})
+}
--- a/audit/internal/platform/runtime.go
+++ b/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,
 	}

--- a/audit/internal/platform/types.go
+++ b/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
--- a/audit/internal/schema/hardware.go
+++ b/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"`
--- a/audit/internal/webui/api.go
+++ b/audit/internal/webui/api.go
@@ -110,7 +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
@@ -127,7 +127,7 @@ func defaultTaskPriority(target string, params taskParams) int {
 		return taskPriorityInstallToRAM
 	case "audit":
 		return taskPriorityAudit
-	case "nvidia-benchmark":
+	case "nvidia-bench-perf", "nvidia-bench-power":
 		return taskPriorityBenchmark
 	case "nvidia-stress", "amd-stress", "memory-stress", "sat-stress", "platform-stress", "nvidia-compute":
 		return taskPriorityBurn
@@ -526,14 +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"`
-				ParallelGPUs       bool     `json:"parallel_gpus"`
-				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"`
@@ -549,14 +549,14 @@ 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,
-				ParallelGPUs:       body.ParallelGPUs,
-				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,
@@ -573,131 +573,142 @@ 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"`
-		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")
+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
 		}
-	}

-	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("nvidia-benchmark", "", "")
-	if strings.TrimSpace(body.DisplayName) != "" {
-		name = body.DisplayName
-	}
-	// Append profile tag.
-	name = fmt.Sprintf("%s · %s", name, profile)
+		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
+			}
+		}

-	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())
+		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
 		}
-		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("benchmark-nvidia"),
-					Name:      stepName,
-					Target:    "nvidia-benchmark",
-					Priority:  defaultTaskPriority("nvidia-benchmark", 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,
-					},
+
+		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)
 				}
-				allTasks = append(allTasks, t)
+				for _, t := range allTasks {
+					globalQueue.enqueue(t)
+				}
+				writeTaskRunResponse(w, allTasks)
+				return
 			}
-			for _, t := range allTasks {
-				globalQueue.enqueue(t)
-			}
-			writeTaskRunResponse(w, allTasks)
+		}
+
+		// 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
 		}
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
 	}
+}

-	// 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("nvidia-benchmark", defaultTaskPriority("nvidia-benchmark", params), time.Now(), params, 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)
+func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
+	h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf").ServeHTTP(w, r)
 }

 func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
@@ -1072,18 +1083,55 @@ 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{
--- a/audit/internal/webui/api_test.go
+++ b/audit/internal/webui/api_test.go
@@ -64,7 +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)
@@ -78,8 +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)
@@ -113,7 +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)
@@ -147,6 +147,50 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	}
 }

+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) {
 	globalQueue.mu.Lock()
 	originalTasks := globalQueue.tasks
@@ -202,7 +246,8 @@ func TestDefaultTaskPriorityOrder(t *testing.T) {
 		defaultTaskPriority("cpu", taskParams{}),
 		defaultTaskPriority("cpu", taskParams{StressMode: true}),
 		defaultTaskPriority("nvidia-stress", taskParams{}),
-		defaultTaskPriority("nvidia-benchmark", taskParams{}),
+		defaultTaskPriority("nvidia-bench-perf", taskParams{}),
+		defaultTaskPriority("nvidia-bench-power", taskParams{}),
 	}
 	want := []int{
 		taskPriorityInstallToRAM,
@@ -211,13 +256,14 @@ func TestDefaultTaskPriorityOrder(t *testing.T) {
 		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]) {
+	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)
 	}
 }
--- a/audit/internal/webui/kmsg_watcher.go
+++ b/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":
--- a/audit/internal/webui/pages.go
+++ b/audit/internal/webui/pages.go
@@ -845,6 +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",
@@ -1939,7 +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>
@@ -1972,21 +1979,25 @@ func renderBenchmark(opts HandlerOptions) string {
        <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>
@@ -2029,21 +2040,24 @@ function benchmarkMode() {

 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');
  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;
+  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). NCCL on final step.';
+    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.' + (selected.length > 1 ? ' NCCL included.' : '');
+    note.textContent = 'Parallel: all ' + selected.length + ' GPU(s) simultaneously. Only the performance benchmark supports this mode.';
  } else {
-    note.textContent = 'Sequential: each GPU benchmarked separately.' + (selected.length > 1 ? ' NCCL included on each.' : '');
+    note.textContent = 'Sequential: each selected GPU benchmarked separately.';
  }
 }

@@ -2117,7 +2131,7 @@ function benchmarkSelectNone() {
  benchmarkUpdateSelectionNote();
 }

-function runNvidiaBenchmark() {
+function runNvidiaBenchmark(kind) {
  const selected = benchmarkSelectedGPUIndices();
  const status = document.getElementById('benchmark-run-status');
  if (!selected.length) {
@@ -2127,21 +2141,26 @@ function runNvidiaBenchmark() {
  if (benchmarkES) { benchmarkES.close(); benchmarkES = null; }
  const mode = benchmarkMode();
  const rampUp = mode === 'ramp-up' && selected.length > 1;
-  const parallelGPUs = mode === 'parallel';
+  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: selected.length > 1,
+    run_nccl: kind === 'performance' && selected.length > 1,
    parallel_gpus: parallelGPUs,
    ramp_up: rampUp,
-    display_name: 'NVIDIA Benchmark'
+    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)
@@ -2195,7 +2214,7 @@ benchmarkLoadGPUs();
 func renderBenchmarkResultsCard(exportDir string) string {
 	maxIdx, runs := loadBenchmarkHistory(exportDir)
 	return renderBenchmarkResultsCardFromRuns(
-		"Benchmark Results",
+		"Perf Results",
 		"Composite score by saved benchmark run and GPU.",
 		"No saved benchmark runs yet.",
 		maxIdx,
@@ -2237,11 +2256,11 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 }

 func loadBenchmarkHistory(exportDir string) (int, []benchmarkHistoryRun) {
-	baseDir := app.DefaultBenchmarkBaseDir
+	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 -1, nil
 	}
@@ -2280,7 +2299,6 @@ func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun)
 	return maxGPUIndex, runs
 }

-
 // ── Burn ──────────────────────────────────────────────────────────────────────

 func renderBurn() string {
@@ -3245,12 +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() {
--- a/audit/internal/webui/server.go
+++ b/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)
--- a/audit/internal/webui/server_test.go
+++ b/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)
 	}
@@ -691,7 +705,7 @@ 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`,
 		`GPU 1`,
@@ -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)
--- a/audit/internal/webui/task_report.go
+++ b/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")
 }

--- a/audit/internal/webui/tasks.go
+++ b/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)",
@@ -628,7 +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
@@ -644,6 +645,19 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			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 {
 			err = fmt.Errorf("app not configured")
--- a/audit/internal/webui/tasks_test.go
+++ b/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)
 	}
@@ -420,7 +420,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	}
 	html := string(body)
 	for _, needle := range []string{
-		`Benchmark Results`,
+		`Perf Results`,
 		`Composite score for this benchmark task.`,
 		`GPU 0`,
 		`1176.25`,
--- a/bible-local/docs/benchmark-clock-calibration.md
+++ b/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.

--- a/iso/builder/bee-gpu-stress.c
+++ b/iso/builder/bee-gpu-stress.c
@@ -642,6 +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",
@@ -760,10 +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);
@@ -776,6 +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;
@@ -944,8 +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;
    }
@@ -1094,17 +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 = &alpha;
+    const void *beta_ptr = &beta;
+    if (profile->desc.compute_type == CUBLAS_COMPUTE_32I) {
+        alpha_ptr = &alpha_i32;
+        beta_ptr = &beta_i32;
+    } else if (profile->desc.compute_type == CUBLAS_COMPUTE_64F) {
+        alpha_ptr = &alpha_f64;
+        beta_ptr = &beta_f64;
+    }
    return check_cublas(profile->desc.name,
                        cublas->cublasLtMatmul(handle,
                                               profile->op_desc,
-                                               &alpha,
+                                               alpha_ptr,
                                               (const void *)(uintptr_t)profile->a_dev,
                                               profile->a_layout,
                                               (const void *)(uintptr_t)profile->b_dev,
                                               profile->b_layout,
-                                               &beta,
+                                               beta_ptr,
                                               (const void *)(uintptr_t)profile->c_dev,
                                               profile->c_layout,
                                               (void *)(uintptr_t)profile->d_dev,
@@ -1359,11 +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]);
@@ -1373,9 +1431,13 @@ int main(int argc, char **argv) {
            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] [--precision fp8|fp16|fp32|fp64|fp4]\n",
+                    "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;
        }
@@ -1436,6 +1498,76 @@ int main(int argc, char **argv) {
    int ok = 0;

 #if HAVE_CUBLASLT_HEADERS
+    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) {
@@ -1454,18 +1586,6 @@ int main(int argc, char **argv) {
        }
    }

-    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;
 }
--- a/iso/overlay/usr/local/bin/bee-gpu-burn
+++ b/iso/overlay/usr/local/bin/bee-gpu-burn
@@ -7,10 +7,12 @@ 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] [--precision fp8|fp16|fp32|fp64|fp4]" >&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
 }

@@ -32,6 +34,8 @@ while [ "$#" -gt 0 ]; do
        --devices) [ "$#" -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
@@ -92,8 +96,12 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
    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}" ${precision_arg} >"${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
Author	SHA1	Message	Date
Michael Chus	a636146dbd	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>	2026-04-14 20:41:17 +03:00
Mikhail Chusavitin	303de2df04	Add slot-aware ramp sequence to bee-bench power	2026-04-14 17:47:40 +03:00
Mikhail Chusavitin	95124d228f	Split bee-bench into perf and power workflows	2026-04-14 17:33:13 +03:00
Mikhail Chusavitin	54338dbae5	Unify live RAM runtime state	2026-04-14 16:18:33 +03:00
Mikhail Chusavitin	2be7ae6d28	Refine NVIDIA benchmark phase timing	2026-04-14 14:12:06 +03:00