Add rsync to initramfs for toram progress output

live-boot already uses rsync --progress when /bin/rsync exists; without it the copy falls back to silent cp -a. Add rsync to the ISO package list and install an initramfs-tools hook (bee-rsync) that copies the rsync binary + shared libs into the initrd via copy_exec. The hook then rebuilds the initramfs so the change takes effect in the ISO's initrd.img. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add toram boot entry and Install to RAM resume support
2026-04-17 23:52:47 +03:00 · 2026-04-17 23:48:56 +03:00 · 2026-04-17 17:59:58 +03:00 · 2026-04-16 11:43:01 +03:00 · 2026-04-16 11:02:30 +03:00 · 2026-04-16 10:54:03 +03:00
23 changed files with 1085 additions and 322 deletions
--- a/audit/internal/app/app.go
+++ b/audit/internal/app/app.go
@@ -146,7 +146,7 @@ type satRunner interface {
 	RunSATStressPack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
 	RunFanStressTest(ctx context.Context, baseDir string, opts platform.FanStressOptions) (string, error)
 	RunPlatformStress(ctx context.Context, baseDir string, opts platform.PlatformStressOptions, logFunc func(string)) (string, error)
-	RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
+	RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
 }

 type runtimeChecker interface {
@@ -744,8 +744,15 @@ func (a *App) RunPlatformStress(ctx context.Context, baseDir string, opts platfo
 	return a.sat.RunPlatformStress(ctx, baseDir, opts, logFunc)
 }

+func (a *App) RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNCCLTests(ctx, baseDir, gpuIndices, logFunc)
+}
+
 func (a *App) RunNCCLTestsResult(ctx context.Context) (ActionResult, error) {
-	path, err := a.sat.RunNCCLTests(ctx, DefaultSATBaseDir, nil)
+	path, err := a.RunNCCLTests(ctx, DefaultSATBaseDir, nil, nil)
 	body := "Results: " + path
 	if err != nil && err != context.Canceled {
 		body += "\nERROR: " + err.Error()
--- a/audit/internal/app/app_test.go
+++ b/audit/internal/app/app_test.go
@@ -128,6 +128,7 @@ type fakeSAT struct {
 	runNvidiaPowerFn          func(string, int, []int) (string, error)
 	runNvidiaPulseFn          func(string, int, []int) (string, error)
 	runNvidiaBandwidthFn      func(string, []int) (string, error)
+	runNCCLFn                 func(string, []int) (string, error)
 	runNvidiaTargetedStressFn func(string, int, []int) (string, error)
 	runMemoryFn               func(string) (string, error)
 	runStorageFn              func(string) (string, error)
@@ -287,10 +288,43 @@ func (f fakeSAT) RunPlatformStress(_ context.Context, _ string, _ platform.Platf
 	return "", nil
 }

-func (f fakeSAT) RunNCCLTests(_ context.Context, _ string, _ func(string)) (string, error) {
+func (f fakeSAT) RunNCCLTests(_ context.Context, baseDir string, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNCCLFn != nil {
+		return f.runNCCLFn(baseDir, gpuIndices)
+	}
 	return "", nil
 }

+func TestRunNCCLTestsPassesSelectedGPUs(t *testing.T) {
+	t.Parallel()
+
+	var gotBaseDir string
+	var gotGPUIndices []int
+	a := &App{
+		sat: fakeSAT{
+			runNCCLFn: func(baseDir string, gpuIndices []int) (string, error) {
+				gotBaseDir = baseDir
+				gotGPUIndices = append([]int(nil), gpuIndices...)
+				return "/tmp/nccl-tests.tar.gz", nil
+			},
+		},
+	}
+
+	path, err := a.RunNCCLTests(context.Background(), "/tmp/sat", []int{3, 1}, nil)
+	if err != nil {
+		t.Fatalf("RunNCCLTests error: %v", err)
+	}
+	if path != "/tmp/nccl-tests.tar.gz" {
+		t.Fatalf("path=%q want %q", path, "/tmp/nccl-tests.tar.gz")
+	}
+	if gotBaseDir != "/tmp/sat" {
+		t.Fatalf("baseDir=%q want %q", gotBaseDir, "/tmp/sat")
+	}
+	if len(gotGPUIndices) != 2 || gotGPUIndices[0] != 3 || gotGPUIndices[1] != 1 {
+		t.Fatalf("gpuIndices=%v want [3 1]", gotGPUIndices)
+	}
+}
+
 func TestNetworkStatusFormatsInterfacesAndRoute(t *testing.T) {
 	t.Parallel()

--- a/audit/internal/platform/benchmark.go
+++ b/audit/internal/platform/benchmark.go
@@ -59,6 +59,9 @@ type benchmarkPowerCalibrationResult struct {
 	// ≥20% while server fans were below 100% duty cycle — a signal that the
 	// cooling system may not be correctly configured for full GPU load.
 	CoolingWarning string
+	// MetricRows holds the telemetry rows from the final (converged) attempt
+	// for this GPU. Used to build per-run gpu-metrics.csv.
+	MetricRows []GPUMetricRow
 }

 type benchmarkBurnProfile struct {
@@ -94,9 +97,13 @@ var (
 )

 // benchmarkPrecisionPhases lists the precision categories run as individual
-// steady-state windows before the combined steady pass.  Order is from lowest
+// steady-state windows before the combined steady pass. Order is from lowest
 // to highest power draw so thermal ramp-up is gradual.
-var benchmarkPrecisionPhases = []string{"int8", "fp8", "fp16", "fp32", "fp64", "fp4"}
+//
+// fp64 and fp4 are intentionally disabled for now: both are currently unstable
+// on the target fleet and can abort the mixed steady stage after the earlier
+// phases already collected useful telemetry.
+var benchmarkPrecisionPhases = []string{"int8", "fp8", "fp16", "fp32"}

 func computeCapabilityCode(raw string) int {
 	raw = strings.TrimSpace(raw)
@@ -124,6 +131,15 @@ func benchmarkSupportedPrecisions(computeCapability string) []string {
 	return out
 }

+func benchmarkPrecisionEnabled(category string) bool {
+	switch category {
+	case "int8", "fp8", "fp16", "fp16_bf16", "fp32", "fp32_tf32":
+		return true
+	default:
+		return false
+	}
+}
+
 func buildBenchmarkSteadyPlan(spec benchmarkProfileSpec, precisions []string, metricStage func(string) string) (planLabels []string, planPhases []benchmarkPlannedPhase, basePhaseSec int, mixedPhaseSec int) {
 	if len(precisions) == 0 {
 		precisions = append([]string(nil), benchmarkPrecisionPhases...)
@@ -514,6 +530,7 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 				appendBenchmarkMetrics(&metricRows, cooldownRows, fmt.Sprintf("gpu-%d-cooldown", idx), &metricTimelineSec, float64(spec.CooldownSec))
 			}

+			applyBenchmarkSteadyFallback(&gpuResult)
 			gpuResult.Scores = scoreBenchmarkGPUResult(gpuResult)
 			gpuResult.DegradationReasons = detectBenchmarkDegradationReasons(gpuResult, result.Normalization.Status)
 			if anomaly := detectPowerAnomaly(metricRows, idx); anomaly != "" {
@@ -1108,6 +1125,7 @@ type benchmarkCoolingSample struct {
 	AvgFanRPM             float64
 	AvgFanDutyCyclePct    float64
 	FanDutyCycleAvailable bool
+	FanDutyCycleEstimated bool
 }

 func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
@@ -1120,6 +1138,7 @@ func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
 		samples[i].FanAvgRPM = fanSample.AvgFanRPM
 		samples[i].FanDutyCyclePct = fanSample.AvgFanDutyCyclePct
 		samples[i].FanDutyCycleAvailable = fanSample.FanDutyCycleAvailable
+		samples[i].FanDutyCycleEstimated = fanSample.FanDutyCycleEstimated
 	}
 	return samples, nil
 }
@@ -1127,11 +1146,12 @@ func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
 func sampleBenchmarkCoolingSample() benchmarkCoolingSample {
 	fans, _ := sampleFanSpeeds()
 	avgRPM, _, _ := fanRPMStats(fans)
-	dutyPct, dutyAvailable := sampleFanDutyCyclePct()
+	dutyPct, dutyAvailable, dutyEstimated := sampleFanDutyCyclePctFromFans(fans)
 	return benchmarkCoolingSample{
 		AvgFanRPM:             avgRPM,
 		AvgFanDutyCyclePct:    dutyPct,
 		FanDutyCycleAvailable: dutyAvailable,
+		FanDutyCycleEstimated: dutyEstimated,
 	}
 }

@@ -1373,44 +1393,91 @@ func summarizeBenchmarkCooling(rows []GPUMetricRow) *BenchmarkCoolingSummary {
 	}
 	var rpmValues []float64
 	var dutyValues []float64
+	var dutyEstimated bool
 	for _, row := range rows {
 		if row.FanAvgRPM > 0 {
 			rpmValues = append(rpmValues, row.FanAvgRPM)
 		}
 		if row.FanDutyCycleAvailable {
 			dutyValues = append(dutyValues, row.FanDutyCyclePct)
+			if row.FanDutyCycleEstimated {
+				dutyEstimated = true
+			}
 		}
 	}
 	if len(rpmValues) == 0 && len(dutyValues) == 0 {
 		return nil
 	}
 	summary := &BenchmarkCoolingSummary{
-		Available: true,
-		AvgFanRPM: benchmarkMean(rpmValues),
+		Available:             true,
+		AvgFanRPM:             benchmarkMean(rpmValues),
+		FanDutyCycleEstimated: dutyEstimated,
 	}
 	if len(dutyValues) > 0 {
 		summary.FanDutyCycleAvailable = true
 		summary.AvgFanDutyCyclePct = benchmarkMean(dutyValues)
 		summary.P95FanDutyCyclePct = benchmarkPercentile(dutyValues, 95)
+		if summary.FanDutyCycleEstimated {
+			summary.Notes = append(summary.Notes, "fan duty cycle is estimated from the highest fan RPM observed since boot; treat it as an approximation, not a direct PWM reading")
+		}
 	} else {
 		summary.Notes = append(summary.Notes, "fan duty cycle unavailable on this host; RPM-only fan telemetry was collected")
 	}
 	return summary
 }

+func benchmarkTelemetryAvailable(summary BenchmarkTelemetrySummary) bool {
+	return summary.Samples > 0 || summary.DurationSec > 0
+}
+
+func benchmarkPrecisionSteadyFallback(phases []BenchmarkPrecisionSteadyPhase) (BenchmarkTelemetrySummary, string, bool) {
+	var (
+		best      BenchmarkTelemetrySummary
+		bestLabel string
+		found     bool
+	)
+	for _, phase := range phases {
+		if !benchmarkTelemetryAvailable(phase.Steady) {
+			continue
+		}
+		if !found ||
+			phase.Steady.DurationSec > best.DurationSec ||
+			(phase.Steady.DurationSec == best.DurationSec && phase.Steady.P95PowerW > best.P95PowerW) {
+			best = phase.Steady
+			bestLabel = phase.Precision
+			found = true
+		}
+	}
+	return best, bestLabel, found
+}
+
+func applyBenchmarkSteadyFallback(gpu *BenchmarkGPUResult) {
+	if gpu == nil || benchmarkTelemetryAvailable(gpu.Steady) {
+		return
+	}
+	if fallback, label, ok := benchmarkPrecisionSteadyFallback(gpu.PrecisionSteady); ok {
+		gpu.Steady = fallback
+		gpu.Notes = append(gpu.Notes,
+			fmt.Sprintf("mixed steady telemetry unavailable; reporting steady-state fallback from %s precision phase", label))
+	}
+}
+
 func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	score := BenchmarkScorecard{}

 	// SyntheticScore: sum of fp32-equivalent TOPS from per-precision phases.
 	// Each precision ran alone with full GPU dedicated — peak capability.
 	for _, p := range gpu.PrecisionSteady {
+		if !benchmarkPrecisionEnabled(p.Precision) {
+			continue
+		}
 		score.SyntheticScore += p.WeightedTeraOpsPerSec
 	}

 	// MixedScore: sum of fp32-equivalent TOPS from the combined phase.
 	// All precisions compete simultaneously — closer to real inference workloads.
 	for _, p := range gpu.PrecisionResults {
-		if p.Supported {
+		if p.Supported && benchmarkPrecisionEnabled(p.Category) {
 			score.MixedScore += p.WeightedTeraOpsPerSec
 		}
 	}
@@ -1441,10 +1508,17 @@ func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	// so CV reflects genuine power regulation, not workload switching).
 	if len(gpu.PrecisionSteady) > 0 {
 		var sum float64
+		var count int
 		for _, p := range gpu.PrecisionSteady {
+			if !benchmarkPrecisionEnabled(p.Precision) {
+				continue
+			}
 			sum += clampScore(100 - p.Steady.PowerCVPct*3)
+			count++
+		}
+		if count > 0 {
+			score.PowerSustainScore = sum / float64(count)
 		}
-		score.PowerSustainScore = sum / float64(len(gpu.PrecisionSteady))
 	} else if gpu.Steady.PowerCVPct > 0 {
 		score.PowerSustainScore = clampScore(100 - gpu.Steady.PowerCVPct*3)
 	}
@@ -2512,6 +2586,7 @@ func runNvidiaBenchmarkParallel(
 	// Score and finalize each GPU.
 	for _, idx := range selected {
 		r := gpuResults[idx]
+		applyBenchmarkSteadyFallback(r)
 		r.Scores = scoreBenchmarkGPUResult(*r)
 		r.DegradationReasons = detectBenchmarkDegradationReasons(*r, result.Normalization.Status)
 		pr := parseResults[idx]
@@ -2694,19 +2769,22 @@ func summarizeCPULoad(samples []float64) *BenchmarkCPULoad {
 	return cl
 }

-// runBenchmarkPowerCalibration runs targeted_power per GPU and actively watches
-// throttle counters. If a GPU starts throttling, the current targeted_power run
-// is canceled immediately, the power limit is reduced, and a fresh full cycle
-// is started again from the beginning. The selected reduced power limit stays
-// active for the main benchmark and is restored by the caller afterwards.
+// runBenchmarkPowerCalibration runs targeted_power for the supplied GPU set and
+// actively watches throttle counters. seedLimits, when provided, are treated as
+// the starting point for this calibration pass rather than as immutable fixed
+// limits. This matters during cumulative ramp-up: once an additional GPU is
+// introduced, every already-active GPU must be revalidated under the new
+// thermal state instead of assuming its previous single-step limit is still
+// valid. The selected reduced power limits stay active for the main benchmark
+// and are restored by the caller afterwards.
 func runBenchmarkPowerCalibration(
 	ctx context.Context,
 	verboseLog, runDir string,
 	gpuIndices []int,
 	infoByIndex map[int]benchmarkGPUInfo,
 	logFunc func(string),
-	fixedLimits map[int]int,
-) (map[int]benchmarkPowerCalibrationResult, []benchmarkRestoreAction) {
+	seedLimits map[int]int,
+) (map[int]benchmarkPowerCalibrationResult, []benchmarkRestoreAction, []GPUMetricRow) {
 	const calibDurationSec = 120
 	const maxDerateW = 150
 	// calibSearchTolerance is the binary-search convergence threshold in watts.
@@ -2720,7 +2798,7 @@ func runBenchmarkPowerCalibration(

 	if _, err := exec.LookPath("dcgmi"); err != nil {
 		logFunc("power calibration: dcgmi not found, skipping (will use default power limit)")
-		return map[int]benchmarkPowerCalibrationResult{}, nil
+		return map[int]benchmarkPowerCalibrationResult{}, nil, nil
 	}
 	if killed := KillTestWorkers(); len(killed) > 0 {
 		for _, p := range killed {
@@ -2739,7 +2817,6 @@ func runBenchmarkPowerCalibration(
 		err  error
 	}

-
 	// gpuCalibState holds per-GPU binary search state during parallel calibration.
 	type gpuCalibState struct {
 		idx            int
@@ -2755,6 +2832,8 @@ func runBenchmarkPowerCalibration(

 	results := make(map[int]benchmarkPowerCalibrationResult, len(gpuIndices))
 	var restore []benchmarkRestoreAction
+	var allCalibRows []GPUMetricRow // accumulated telemetry across all attempts
+	var calibCursor float64

 	// Initialise per-GPU state.
 	states := make([]*gpuCalibState, 0, len(gpuIndices))
@@ -2796,19 +2875,20 @@ func runBenchmarkPowerCalibration(
 			hi:             appliedLimitW + 1, // not yet tested, not yet confirmed unstable
 			calib:          benchmarkPowerCalibrationResult{AppliedPowerLimitW: float64(appliedLimitW)},
 		}
-		if fixedLimits != nil {
-			if fixedW, ok := fixedLimits[idx]; ok {
-				// This GPU's limit was established in a prior ramp step and must
-				// remain unchanged. Apply it immediately and skip the binary search.
-				if canDerate && fixedW > 0 {
-					_ = setBenchmarkPowerLimit(ctx, verboseLog, idx, fixedW)
+		if seedLimits != nil {
+			if seedW, ok := seedLimits[idx]; ok && seedW > 0 {
+				// A previously validated limit is only a starting point. Re-run
+				// targeted_power under the current multi-GPU thermal load and derate
+				// again if this step shows new throttling.
+				if canDerate {
+					_ = setBenchmarkPowerLimit(ctx, verboseLog, idx, seedW)
 				}
-				s.appliedLimitW = fixedW
-				s.calib.AppliedPowerLimitW = float64(fixedW)
-				s.calib.Completed = true
-				s.converged = true
+				s.appliedLimitW = seedW
+				s.hi = seedW + 1
+				s.calib.AppliedPowerLimitW = float64(seedW)
+				s.calib.Derated = seedW < s.originalLimitW
 				s.calib.Notes = append(s.calib.Notes,
-					fmt.Sprintf("fixed limit: %d W (held from prior ramp step)", fixedW))
+					fmt.Sprintf("seed limit: %d W (revalidating under current thermal load)", seedW))
 			}
 		}
 		states = append(states, s)
@@ -2906,6 +2986,8 @@ calibDone:
 		ticker.Stop()
 		cancelAttempt()
 		_ = os.WriteFile(filepath.Join(runDir, logName), ar.out, 0644)
+		// Accumulate telemetry rows with attempt stage label.
+		appendBenchmarkMetrics(&allCalibRows, ar.rows, fmt.Sprintf("attempt-%d", sharedAttempt), &calibCursor, float64(calibDurationSec))

 		// Resource busy: retry with exponential back-off (shared — one DCGM session).
 		if ar.err != nil && isDCGMResourceBusy(ar.err) {
@@ -2990,6 +3072,7 @@ calibDone:
 						}
 					}
 				}
+				s.calib.MetricRows = filterRowsByGPU(ar.rows, s.idx)
 				s.converged = true
 				continue
 			}
@@ -3028,6 +3111,7 @@ calibDone:
 				} else {
 					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
 				}
+				s.calib.MetricRows = filterRowsByGPU(ar.rows, s.idx)
 				s.converged = true
 				continue
 			}
@@ -3065,7 +3149,8 @@ calibDone:
 			results[s.idx] = s.calib
 		}
 	}
-	return results, restore
+	writeBenchmarkMetricsFiles(runDir, allCalibRows)
+	return results, restore, allCalibRows
 }

 // isDCGMResourceBusy returns true when dcgmi exits with DCGM_ST_IN_USE (222),
@@ -3091,7 +3176,6 @@ func powerBenchDurationSec(profile string) int {
 	}
 }

-
 func cloneBenchmarkGPUInfoMap(src map[int]benchmarkGPUInfo) map[int]benchmarkGPUInfo {
 	out := make(map[int]benchmarkGPUInfo, len(src))
 	for k, v := range src {
@@ -3107,7 +3191,42 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 	fmt.Fprintf(&b, "**Profile:** %s  \n", result.BenchmarkProfile)
 	fmt.Fprintf(&b, "**Generated:** %s  \n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
 	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
-	fmt.Fprintf(&b, "**Platform max TDP:** %.0f W  \n\n", result.PlatformMaxTDPW)
+	fmt.Fprintf(&b, "**Platform max TDP (GPU-reported):** %.0f W  \n", result.PlatformMaxTDPW)
+	if sp := result.ServerPower; sp != nil && sp.Available {
+		fmt.Fprintf(&b, "**Server power delta (IPMI):** %.0f W  \n", sp.DeltaW)
+		fmt.Fprintf(&b, "**Reporting ratio (IPMI Δ / GPU sum):** %.2f  \n", sp.ReportingRatio)
+	}
+	b.WriteString("\n")
+	// Server power comparison table.
+	if sp := result.ServerPower; sp != nil {
+		b.WriteString("## Server vs GPU Power Comparison\n\n")
+		b.WriteString("| Metric | Value |\n")
+		b.WriteString("|--------|-------|\n")
+		fmt.Fprintf(&b, "| GPU stable limits sum (nvidia-smi) | %.0f W |\n", result.PlatformMaxTDPW)
+		if sp.Available {
+			fmt.Fprintf(&b, "| Server idle power (IPMI) | %.0f W |\n", sp.IdleW)
+			fmt.Fprintf(&b, "| Server loaded power (IPMI) | %.0f W |\n", sp.LoadedW)
+			fmt.Fprintf(&b, "| Server Δ power (loaded − idle) | %.0f W |\n", sp.DeltaW)
+			ratio := sp.ReportingRatio
+			ratioNote := ""
+			switch {
+			case ratio >= 0.9:
+				ratioNote = "✓ GPU telemetry matches server power"
+			case ratio >= 0.75:
+				ratioNote = "⚠ minor discrepancy — GPU may slightly over-report TDP"
+			default:
+				ratioNote = "✗ significant discrepancy — GPU over-reports TDP vs wall power"
+			}
+			fmt.Fprintf(&b, "| Reporting ratio (IPMI Δ / GPU sum) | %.2f — %s |\n", ratio, ratioNote)
+		} else {
+			b.WriteString("| IPMI availability | not available — IPMI not supported or ipmitool not found |\n")
+		}
+		for _, note := range sp.Notes {
+			fmt.Fprintf(&b, "\n> %s\n", note)
+		}
+		b.WriteString("\n")
+	}
+
 	if len(result.Findings) > 0 {
 		b.WriteString("## Summary\n\n")
 		for _, finding := range result.Findings {
@@ -3121,21 +3240,25 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 	}
 	if len(result.RampSteps) > 0 {
 		b.WriteString("## Ramp Sequence\n\n")
-		b.WriteString("| Step | New GPU | Stable Limit | Total Observed | Derated | Status |\n")
-		b.WriteString("|------|---------|--------------|----------------|---------|--------|\n")
+		b.WriteString("| Step | New GPU | Stable Limit | Total Observed | Server Δ (IPMI) | Derated | Status |\n")
+		b.WriteString("|------|---------|--------------|----------------|-----------------|---------|--------|\n")
 		for _, step := range result.RampSteps {
 			derated := "-"
 			if step.Derated {
 				derated = "⚠ yes"
 			}
-			fmt.Fprintf(&b, "| %d | GPU %d | %.0f W | %.0f W | %s | %s |\n",
-				step.StepIndex, step.NewGPUIndex, step.NewGPUStableLimitW, step.TotalObservedPowerW, derated, step.Status)
+			serverDelta := "-"
+			if step.ServerDeltaW > 0 {
+				serverDelta = fmt.Sprintf("%.0f W", step.ServerDeltaW)
+			}
+			fmt.Fprintf(&b, "| %d | GPU %d | %.0f W | %.0f W | %s | %s | %s |\n",
+				step.StepIndex, step.NewGPUIndex, step.NewGPUStableLimitW, step.TotalObservedPowerW, serverDelta, derated, step.Status)
 		}
 		b.WriteString("\n")
 	}
 	b.WriteString("## Per-Slot Results\n\n")
-	b.WriteString("| GPU | Status | Single-card Limit | Stable Limit | Temp | Attempts |\n")
-	b.WriteString("|-----|--------|-------------------|--------------|------|----------|\n")
+	b.WriteString("| GPU | Status | Single-card Limit | Stable Limit | Server Δ (IPMI) | Temp | Attempts |\n")
+	b.WriteString("|-----|--------|-------------------|--------------|-----------------|------|----------|\n")
 	for _, gpu := range result.GPUs {
 		stableLimit := "-"
 		if gpu.StablePowerLimitW > 0 {
@@ -3145,8 +3268,12 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				stableLimit = fmt.Sprintf("%.0f W", gpu.StablePowerLimitW)
 			}
 		}
-		fmt.Fprintf(&b, "| GPU %d | %s | %.0f W | %s | %.1f C | %d |\n",
-			gpu.Index, gpu.Status, gpu.AppliedPowerLimitW, stableLimit, gpu.MaxObservedTempC, gpu.CalibrationAttempts)
+		serverDelta := "-"
+		if gpu.ServerDeltaW > 0 {
+			serverDelta = fmt.Sprintf("%.0f W", gpu.ServerDeltaW)
+		}
+		fmt.Fprintf(&b, "| GPU %d | %s | %.0f W | %s | %s | %.1f C | %d |\n",
+			gpu.Index, gpu.Status, gpu.AppliedPowerLimitW, stableLimit, serverDelta, gpu.MaxObservedTempC, gpu.CalibrationAttempts)
 	}
 	b.WriteString("\n")
 	for _, gpu := range result.GPUs {
@@ -3175,11 +3302,25 @@ func renderPowerBenchSummary(result NvidiaPowerBenchResult) string {
 		fmt.Fprintf(&b, "ramp_step_%d_new_gpu=%d\n", step.StepIndex, step.NewGPUIndex)
 		fmt.Fprintf(&b, "ramp_step_%d_stable_limit_w=%.0f\n", step.StepIndex, step.NewGPUStableLimitW)
 		fmt.Fprintf(&b, "ramp_step_%d_total_power_w=%.0f\n", step.StepIndex, step.TotalObservedPowerW)
+		if step.ServerLoadedW > 0 {
+			fmt.Fprintf(&b, "ramp_step_%d_server_loaded_w=%.0f\n", step.StepIndex, step.ServerLoadedW)
+			fmt.Fprintf(&b, "ramp_step_%d_server_delta_w=%.0f\n", step.StepIndex, step.ServerDeltaW)
+		}
 	}
 	for _, gpu := range result.GPUs {
 		if gpu.StablePowerLimitW > 0 {
 			fmt.Fprintf(&b, "gpu_%d_stable_limit_w=%.0f\n", gpu.Index, gpu.StablePowerLimitW)
 		}
+		if gpu.ServerLoadedW > 0 {
+			fmt.Fprintf(&b, "gpu_%d_server_loaded_w=%.0f\n", gpu.Index, gpu.ServerLoadedW)
+			fmt.Fprintf(&b, "gpu_%d_server_delta_w=%.0f\n", gpu.Index, gpu.ServerDeltaW)
+		}
+	}
+	if sp := result.ServerPower; sp != nil && sp.Available {
+		fmt.Fprintf(&b, "server_idle_w=%.0f\n", sp.IdleW)
+		fmt.Fprintf(&b, "server_loaded_w=%.0f\n", sp.LoadedW)
+		fmt.Fprintf(&b, "server_delta_w=%.0f\n", sp.DeltaW)
+		fmt.Fprintf(&b, "server_reporting_ratio=%.2f\n", sp.ReportingRatio)
 	}
 	return b.String()
 }
@@ -3212,6 +3353,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	if infoErr != nil {
 		return "", infoErr
 	}
+	// Capture full nvidia-smi -q snapshot at the start of the run.
+	if out, err := runSATCommandCtx(ctx, verboseLog, "00-nvidia-smi-q.log", []string{"nvidia-smi", "-q"}, nil, nil); err == nil {
+		_ = os.WriteFile(filepath.Join(runDir, "00-nvidia-smi-q.log"), out, 0644)
+	}
 	hostname, _ := os.Hostname()
 	result := NvidiaPowerBenchResult{
 		BenchmarkVersion:   benchmarkVersion,
@@ -3224,16 +3369,44 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	}
 	durationSec := powerBenchDurationSec(opts.Profile)
 	_ = durationSec
+
+	// Sample IPMI idle power before any GPU load.
+	var serverIdleW float64
+	var serverIdleOK bool
+	if w, ok := sampleIPMIPowerSeries(ctx, 10); ok {
+		serverIdleW = w
+		serverIdleOK = true
+		logFunc(fmt.Sprintf("server idle power (IPMI): %.0f W", w))
+	}
+
 	// Phase 1: calibrate each GPU individually (sequentially, one at a time) to
 	// establish a true single-card power baseline unaffected by neighbour heat.
 	calibByIndex := make(map[int]benchmarkPowerCalibrationResult, len(selected))
+	singleIPMILoadedW := make(map[int]float64, len(selected))
 	var allRestoreActions []benchmarkRestoreAction
+	// allPowerRows accumulates telemetry from all phases for the top-level gpu-metrics.csv.
+	var allPowerRows []GPUMetricRow
+	var powerCursor float64
 	for _, idx := range selected {
 		singleDir := filepath.Join(runDir, fmt.Sprintf("single-%02d", idx))
 		_ = os.MkdirAll(singleDir, 0755)
 		singleInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
 		logFunc(fmt.Sprintf("power calibration: GPU %d single-card baseline", idx))
-		c, restore := runBenchmarkPowerCalibration(ctx, verboseLog, singleDir, []int{idx}, singleInfo, logFunc, nil)
+		ipmiSingleCtx, ipmiSingleCancel := context.WithCancel(ctx)
+		ipmiSingleDone := make(chan float64, 1)
+		go func() {
+			defer close(ipmiSingleDone)
+			if w, ok := sampleIPMIPowerSeries(ipmiSingleCtx, 3600); ok {
+				ipmiSingleDone <- w
+			}
+		}()
+		c, restore, singleRows := runBenchmarkPowerCalibration(ctx, verboseLog, singleDir, []int{idx}, singleInfo, logFunc, nil)
+		appendBenchmarkMetrics(&allPowerRows, singleRows, fmt.Sprintf("single-gpu-%d", idx), &powerCursor, 0)
+		ipmiSingleCancel()
+		if w, ok := <-ipmiSingleDone; ok {
+			singleIPMILoadedW[idx] = w
+			logFunc(fmt.Sprintf("power calibration: GPU %d single-card IPMI loaded: %.0f W", idx, w))
+		}
 		allRestoreActions = append(allRestoreActions, restore...)
 		if r, ok := c[idx]; ok {
 			calibByIndex[idx] = r
@@ -3258,7 +3431,7 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 				result.OverallStatus = "PARTIAL"
 			}
 		}
-		gpus = append(gpus, NvidiaPowerBenchGPU{
+		gpu := NvidiaPowerBenchGPU{
 			Index:               idx,
 			Name:                info.Name,
 			BusID:               info.BusID,
@@ -3271,7 +3444,16 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			Status:              status,
 			Notes:               append([]string(nil), calib.Notes...),
 			CoolingWarning:      calib.CoolingWarning,
-		})
+		}
+		if w, ok := singleIPMILoadedW[idx]; ok && serverIdleOK && w > 0 {
+			gpu.ServerLoadedW = w
+			gpu.ServerDeltaW = w - serverIdleW
+		}
+		if len(calib.MetricRows) > 0 {
+			t := summarizeBenchmarkTelemetry(calib.MetricRows)
+			gpu.Telemetry = &t
+		}
+		gpus = append(gpus, gpu)
 	}
 	sort.Slice(gpus, func(i, j int) bool {
 		if gpus[i].MaxObservedPowerW != gpus[j].MaxObservedPowerW {
@@ -3320,20 +3502,30 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	// stableLimits accumulates GPU index → fixed stable limit (W) across steps.
 	stableLimits := make(map[int]int, len(result.RecommendedSlotOrder))

+	// serverLoadedW tracks the IPMI server power from the final ramp step
+	// (all GPUs simultaneously loaded). Earlier steps' values are stored
+	// per-step in NvidiaPowerBenchStep.ServerLoadedW.
+	var serverLoadedW float64
+	var serverLoadedOK bool
+
 	// Step 1: reuse single-card calibration result directly.
 	if len(result.RecommendedSlotOrder) > 0 {
 		firstIdx := result.RecommendedSlotOrder[0]
 		firstCalib := calibByIndex[firstIdx]
 		stableLimits[firstIdx] = int(math.Round(firstCalib.AppliedPowerLimitW))
 		ramp := NvidiaPowerBenchStep{
-			StepIndex:         1,
-			GPUIndices:        []int{firstIdx},
-			NewGPUIndex:       firstIdx,
-			NewGPUStableLimitW: firstCalib.AppliedPowerLimitW,
+			StepIndex:           1,
+			GPUIndices:          []int{firstIdx},
+			NewGPUIndex:         firstIdx,
+			NewGPUStableLimitW:  firstCalib.AppliedPowerLimitW,
 			TotalObservedPowerW: firstCalib.Summary.P95PowerW,
 			AvgObservedPowerW:   firstCalib.Summary.P95PowerW,
-			Derated:           firstCalib.Derated,
-			Status:            "OK",
+			Derated:             firstCalib.Derated,
+			Status:              "OK",
+		}
+		if w, ok := singleIPMILoadedW[firstIdx]; ok && serverIdleOK && w > 0 {
+			ramp.ServerLoadedW = w
+			ramp.ServerDeltaW = w - serverIdleW
 		}
 		if !firstCalib.Completed {
 			ramp.Status = "FAILED"
@@ -3351,8 +3543,9 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			len(result.RecommendedSlotOrder), firstIdx, firstCalib.AppliedPowerLimitW))
 	}

-	// Steps 2..N: each step fixes previously calibrated GPUs and searches only
-	// the new GPU's stable limit in the combined thermal environment.
+	// Steps 2..N: each step revalidates every already-active GPU under the new
+	// cumulative thermal environment and also calibrates the newly introduced
+	// GPU. Previously found limits are used only as seeds for the search.
 	for stepNum := 1; stepNum < len(result.RecommendedSlotOrder); stepNum++ {
 		step := stepNum + 1
 		subset := append([]int(nil), result.RecommendedSlotOrder[:step]...)
@@ -3360,17 +3553,46 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		stepDir := filepath.Join(runDir, fmt.Sprintf("step-%02d", step))
 		_ = os.MkdirAll(stepDir, 0755)

-		// All previously calibrated GPUs are fixed at their stable limits.
-		fixedForStep := make(map[int]int, len(stableLimits))
-		for k, v := range stableLimits {
-			fixedForStep[k] = v
+		// Reuse the latest stable limits as starting points, but re-check every
+		// active GPU in this hotter configuration. For the newly introduced GPU,
+		// seed from its single-card calibration so we do not restart from the
+		// default TDP when a prior derated limit is already known.
+		seedForStep := make(map[int]int, len(subset))
+		for _, idx := range subset {
+			if lim, ok := stableLimits[idx]; ok && lim > 0 {
+				seedForStep[idx] = lim
+				continue
+			}
+			if base, ok := calibByIndex[idx]; ok {
+				lim := int(math.Round(base.AppliedPowerLimitW))
+				if lim > 0 {
+					seedForStep[idx] = lim
+				}
+			}
 		}

-		logFunc(fmt.Sprintf("power ramp: step %d/%d — calibrating GPU %d with %d fixed GPU(s)",
-			step, len(result.RecommendedSlotOrder), newGPUIdx, len(fixedForStep)))
+		logFunc(fmt.Sprintf("power ramp: step %d/%d — revalidating %d active GPU(s) including new GPU %d",
+			step, len(result.RecommendedSlotOrder), len(subset), newGPUIdx))

 		stepInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
-		stepCalib, stepRestore := runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc, fixedForStep)
+		ipmiStepCtx, ipmiStepCancel := context.WithCancel(ctx)
+		ipmiStepDone := make(chan float64, 1)
+		go func() {
+			defer close(ipmiStepDone)
+			if w, ok := sampleIPMIPowerSeries(ipmiStepCtx, 3600); ok {
+				ipmiStepDone <- w
+			}
+		}()
+		stepCalib, stepRestore, stepRows := runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc, seedForStep)
+		appendBenchmarkMetrics(&allPowerRows, stepRows, fmt.Sprintf("ramp-step-%d", step), &powerCursor, 0)
+		ipmiStepCancel()
+		var stepIPMILoadedW float64
+		var stepIPMIOK bool
+		if w, ok := <-ipmiStepDone; ok {
+			stepIPMILoadedW = w
+			stepIPMIOK = true
+			logFunc(fmt.Sprintf("power ramp: step %d IPMI loaded: %.0f W", step, w))
+		}
 		// Accumulate restore actions; they all run in the outer defer.
 		allRestoreActions = append(allRestoreActions, stepRestore...)

@@ -3391,26 +3613,56 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			ramp.AvgObservedPowerW = ramp.TotalObservedPowerW / float64(len(subset))
 		}

-		// Determine stable limit for the new GPU.
-		if c, ok := stepCalib[newGPUIdx]; ok && c.Completed {
-			stableLimits[newGPUIdx] = int(math.Round(c.AppliedPowerLimitW))
-			ramp.NewGPUStableLimitW = c.AppliedPowerLimitW
-			ramp.Derated = c.Derated
+		for _, idx := range subset {
+			c, ok := stepCalib[idx]
+			if !ok || !c.Completed {
+				fallback := 0
+				if lim, ok := stableLimits[idx]; ok && lim > 0 {
+					fallback = lim
+				} else if fb, ok := calibByIndex[idx]; ok {
+					fallback = int(math.Round(fb.AppliedPowerLimitW))
+				}
+				if fallback > 0 {
+					stableLimits[idx] = fallback
+				}
+				ramp.Status = "FAILED"
+				ramp.Notes = append(ramp.Notes,
+					fmt.Sprintf("GPU %d did not complete targeted_power in ramp step %d; keeping previous stable limit %d W", idx, step, fallback))
+				result.OverallStatus = "PARTIAL"
+				continue
+			}
+
+			prevLimit, hadPrev := stableLimits[idx]
+			newLimit := int(math.Round(c.AppliedPowerLimitW))
+			stableLimits[idx] = newLimit
+			if idx == newGPUIdx {
+				ramp.NewGPUStableLimitW = c.AppliedPowerLimitW
+				ramp.Derated = c.Derated
+			}
 			if c.Derated {
 				ramp.Status = "PARTIAL"
 				if result.OverallStatus == "OK" {
 					result.OverallStatus = "PARTIAL"
 				}
-				result.Findings = append(result.Findings, fmt.Sprintf("Ramp step %d (GPU %d) required derating to %.0f W under combined thermal load.", step, newGPUIdx, c.AppliedPowerLimitW))
 			}
-		} else {
-			// Calibration failed — fall back to single-card limit.
-			fb := calibByIndex[newGPUIdx]
-			stableLimits[newGPUIdx] = int(math.Round(fb.AppliedPowerLimitW))
-			ramp.NewGPUStableLimitW = fb.AppliedPowerLimitW
-			ramp.Status = "FAILED"
-			ramp.Notes = append(ramp.Notes, fmt.Sprintf("GPU %d did not complete targeted_power in ramp step %d; using single-card limit %.0f W", newGPUIdx, step, fb.AppliedPowerLimitW))
-			result.OverallStatus = "PARTIAL"
+			if hadPrev && newLimit < prevLimit {
+				ramp.Notes = append(ramp.Notes,
+					fmt.Sprintf("GPU %d was re-derated from %d W to %d W under combined thermal load.", idx, prevLimit, newLimit))
+			}
+		}
+
+		if c, ok := stepCalib[newGPUIdx]; ok && c.Completed && c.Derated {
+			result.Findings = append(result.Findings, fmt.Sprintf("Ramp step %d (GPU %d) required derating to %.0f W under combined thermal load.", step, newGPUIdx, c.AppliedPowerLimitW))
+		}
+
+		if stepIPMIOK && serverIdleOK && stepIPMILoadedW > 0 {
+			ramp.ServerLoadedW = stepIPMILoadedW
+			ramp.ServerDeltaW = stepIPMILoadedW - serverIdleW
+			// The last step has all GPUs loaded — use it as the top-level loaded_w.
+			if step == len(result.RecommendedSlotOrder) {
+				serverLoadedW = stepIPMILoadedW
+				serverLoadedOK = true
+			}
 		}

 		result.RampSteps = append(result.RampSteps, ramp)
@@ -3421,6 +3673,14 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		if lim, ok := stableLimits[result.GPUs[i].Index]; ok {
 			result.GPUs[i].StablePowerLimitW = float64(lim)
 		}
+		if result.GPUs[i].StablePowerLimitW > 0 && result.GPUs[i].AppliedPowerLimitW > 0 &&
+			result.GPUs[i].StablePowerLimitW < result.GPUs[i].AppliedPowerLimitW {
+			result.GPUs[i].Derated = true
+			result.Findings = append(result.Findings, fmt.Sprintf(
+				"GPU %d required additional derating from %.0f W (single-card) to %.0f W under full-system thermal load.",
+				result.GPUs[i].Index, result.GPUs[i].AppliedPowerLimitW, result.GPUs[i].StablePowerLimitW,
+			))
+		}
 	}

 	// PlatformMaxTDPW = sum of all stable limits — the actual sustained power
@@ -3428,6 +3688,15 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	for _, lim := range stableLimits {
 		result.PlatformMaxTDPW += float64(lim)
 	}
+
+	// Characterize server power from IPMI idle/loaded samples.
+	// GPUReportedSumW = PlatformMaxTDPW (sum of stable GPU limits, nvidia-smi).
+	// ReportingRatio = IPMI_delta / GPU_reported_sum:
+	//   ~1.0 → GPU telemetry matches wall power; <0.75 → GPU over-reports its TDP.
+	_ = serverIdleOK // used implicitly via characterizeServerPower
+	result.ServerPower = characterizeServerPower(serverIdleW, serverLoadedW, result.PlatformMaxTDPW, serverIdleOK && serverLoadedOK)
+	// Write top-level gpu-metrics.csv/.html aggregating all phases.
+	writeBenchmarkMetricsFiles(runDir, allPowerRows)
 	resultJSON, err := json.MarshalIndent(result, "", "  ")
 	if err != nil {
 		return "", fmt.Errorf("marshal power result: %w", err)
--- a/audit/internal/platform/benchmark_report.go
+++ b/audit/internal/platform/benchmark_report.go
@@ -261,14 +261,18 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		b.WriteString("\n")

 		// Steady-state telemetry
-		fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
-		b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
-		fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
-		fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
-		fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
-		fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
-		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
-		b.WriteString("\n")
+		if benchmarkTelemetryAvailable(gpu.Steady) {
+			fmt.Fprintf(&b, "**Steady-state telemetry** (%ds):\n\n", int(gpu.Steady.DurationSec))
+			b.WriteString("| | Avg | P95 |\n|---|---|---|\n")
+			fmt.Fprintf(&b, "| Power | %.1f W | %.1f W |\n", gpu.Steady.AvgPowerW, gpu.Steady.P95PowerW)
+			fmt.Fprintf(&b, "| Temperature | %.1f °C | %.1f °C |\n", gpu.Steady.AvgTempC, gpu.Steady.P95TempC)
+			fmt.Fprintf(&b, "| GPU clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgGraphicsClockMHz, gpu.Steady.P95GraphicsClockMHz)
+			fmt.Fprintf(&b, "| Memory clock | %.0f MHz | %.0f MHz |\n", gpu.Steady.AvgMemoryClockMHz, gpu.Steady.P95MemoryClockMHz)
+			fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
+			b.WriteString("\n")
+		} else {
+			b.WriteString("**Steady-state telemetry:** unavailable\n\n")
+		}

 		// Per-precision stability phases.
 		if len(gpu.PrecisionSteady) > 0 {
--- a/audit/internal/platform/benchmark_test.go
+++ b/audit/internal/platform/benchmark_test.go
@@ -49,8 +49,8 @@ func TestBuildBenchmarkSteadyPlanStandard(t *testing.T) {
 		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 len(labels) != 5 || len(phases) != 5 {
+		t.Fatalf("labels=%d phases=%d want 5", len(labels), len(phases))
 	}
 	if basePhaseSec != 60 {
 		t.Fatalf("basePhaseSec=%d want 60", basePhaseSec)
@@ -61,7 +61,7 @@ func TestBuildBenchmarkSteadyPlanStandard(t *testing.T) {
 	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" {
+	if benchmarkPlanDurationsCSV(phases) != "60,60,60,60,300" {
 		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
 	}
 }
@@ -80,7 +80,7 @@ func TestBuildBenchmarkSteadyPlanStability(t *testing.T) {
 	if mixedPhaseSec != 3600 {
 		t.Fatalf("mixedPhaseSec=%d want 3600", mixedPhaseSec)
 	}
-	if benchmarkPlanDurationsCSV(phases) != "300,300,300,300,300,300,3600" {
+	if benchmarkPlanDurationsCSV(phases) != "300,300,300,300,3600" {
 		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
 	}
 }
@@ -99,7 +99,7 @@ func TestBuildBenchmarkSteadyPlanOvernight(t *testing.T) {
 	if mixedPhaseSec != 14400 {
 		t.Fatalf("mixedPhaseSec=%d want 14400", mixedPhaseSec)
 	}
-	if benchmarkPlanDurationsCSV(phases) != "3600,3600,3600,3600,3600,3600,14400" {
+	if benchmarkPlanDurationsCSV(phases) != "3600,3600,3600,3600,14400" {
 		t.Fatalf("durations=%q", benchmarkPlanDurationsCSV(phases))
 	}
 }
@@ -133,10 +133,10 @@ func TestSplitBenchmarkRowsByPlannedPhaseUsesPhaseDurations(t *testing.T) {
 func TestBenchmarkSupportedPrecisionsSkipsFP4BeforeBlackwell(t *testing.T) {
 	t.Parallel()

-	if got := benchmarkSupportedPrecisions("9.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32,fp64" {
+	if got := benchmarkSupportedPrecisions("9.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32" {
 		t.Fatalf("supported=%v", got)
 	}
-	if got := benchmarkSupportedPrecisions("10.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32,fp64,fp4" {
+	if got := benchmarkSupportedPrecisions("10.0"); strings.Join(got, ",") != "int8,fp8,fp16,fp32" {
 		t.Fatalf("supported=%v", got)
 	}
 }
@@ -314,6 +314,30 @@ func TestRenderBenchmarkReportListsUnifiedArtifacts(t *testing.T) {
 	}
 }

+func TestScoreBenchmarkGPUIgnoresDisabledPrecisions(t *testing.T) {
+	t.Parallel()
+
+	score := scoreBenchmarkGPUResult(BenchmarkGPUResult{
+		PrecisionSteady: []BenchmarkPrecisionSteadyPhase{
+			{Precision: "fp16", WeightedTeraOpsPerSec: 100},
+			{Precision: "fp64", WeightedTeraOpsPerSec: 999},
+			{Precision: "fp4", WeightedTeraOpsPerSec: 999},
+		},
+		PrecisionResults: []BenchmarkPrecisionResult{
+			{Category: "fp32_tf32", Supported: true, WeightedTeraOpsPerSec: 50},
+			{Category: "fp64", Supported: true, WeightedTeraOpsPerSec: 999},
+			{Category: "fp4", Supported: true, WeightedTeraOpsPerSec: 999},
+		},
+	})
+
+	if score.SyntheticScore != 100 {
+		t.Fatalf("SyntheticScore=%f want 100", score.SyntheticScore)
+	}
+	if score.MixedScore != 50 {
+		t.Fatalf("MixedScore=%f want 50", score.MixedScore)
+	}
+}
+
 func TestEnrichGPUInfoWithMaxClocks(t *testing.T) {
 	t.Parallel()

--- a/audit/internal/platform/benchmark_types.go
+++ b/audit/internal/platform/benchmark_types.go
@@ -31,6 +31,7 @@ type BenchmarkCoolingSummary struct {
 	Available             bool     `json:"available"`
 	AvgFanRPM             float64  `json:"avg_fan_rpm,omitempty"`
 	FanDutyCycleAvailable bool     `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool     `json:"fan_duty_cycle_estimated,omitempty"`
 	AvgFanDutyCyclePct    float64  `json:"avg_fan_duty_cycle_pct,omitempty"`
 	P95FanDutyCyclePct    float64  `json:"p95_fan_duty_cycle_pct,omitempty"`
 	Notes                 []string `json:"notes,omitempty"`
@@ -55,32 +56,32 @@ type NvidiaBenchmarkOptions struct {
 }

 type NvidiaBenchmarkResult struct {
-	BenchmarkVersion   string                       `json:"benchmark_version"`
-	GeneratedAt        time.Time                    `json:"generated_at"`
-	Hostname           string                       `json:"hostname,omitempty"`
-	ServerModel        string                       `json:"server_model,omitempty"`
-	BenchmarkProfile   string                       `json:"benchmark_profile"`
-	ParallelGPUs       bool                         `json:"parallel_gpus,omitempty"`
-	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"`
+	BenchmarkVersion string    `json:"benchmark_version"`
+	GeneratedAt      time.Time `json:"generated_at"`
+	Hostname         string    `json:"hostname,omitempty"`
+	ServerModel      string    `json:"server_model,omitempty"`
+	BenchmarkProfile string    `json:"benchmark_profile"`
+	ParallelGPUs     bool      `json:"parallel_gpus,omitempty"`
+	RampStep         int       `json:"ramp_step,omitempty"`
+	RampTotal        int       `json:"ramp_total,omitempty"`
+	RampRunID        string    `json:"ramp_run_id,omitempty"`
+	ScalabilityScore float64   `json:"scalability_score,omitempty"`
 	// PlatformPowerScore is the mean compute scalability across ramp steps 2..N.
 	// 100% = each added GPU contributes exactly its single-card throughput.
 	// < 100% = throughput loss due to thermal throttle, power limits, or contention.
-	PlatformPowerScore   float64                    `json:"platform_power_score,omitempty"`
-	PerformanceRampSteps []NvidiaPerformanceRampStep `json:"performance_ramp_steps,omitempty"`
-	OverallStatus      string                       `json:"overall_status"`
-	SelectedGPUIndices []int                        `json:"selected_gpu_indices"`
-	Findings           []string                     `json:"findings,omitempty"`
-	Warnings           []string                     `json:"warnings,omitempty"`
-	Normalization      BenchmarkNormalization       `json:"normalization"`
-	HostConfig         *BenchmarkHostConfig         `json:"host_config,omitempty"`
-	CPULoad            *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
-	Cooling            *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
-	GPUs               []BenchmarkGPUResult         `json:"gpus"`
-	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
-	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`
+	PlatformPowerScore   float64                      `json:"platform_power_score,omitempty"`
+	PerformanceRampSteps []NvidiaPerformanceRampStep  `json:"performance_ramp_steps,omitempty"`
+	OverallStatus        string                       `json:"overall_status"`
+	SelectedGPUIndices   []int                        `json:"selected_gpu_indices"`
+	Findings             []string                     `json:"findings,omitempty"`
+	Warnings             []string                     `json:"warnings,omitempty"`
+	Normalization        BenchmarkNormalization       `json:"normalization"`
+	HostConfig           *BenchmarkHostConfig         `json:"host_config,omitempty"`
+	CPULoad              *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
+	Cooling              *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
+	GPUs                 []BenchmarkGPUResult         `json:"gpus"`
+	Interconnect         *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
+	ServerPower          *BenchmarkServerPower        `json:"server_power,omitempty"`
 }

 type BenchmarkNormalization struct {
@@ -223,8 +224,8 @@ type BenchmarkScorecard struct {

 	// Throttle breakdown — percentage of steady-state time in each throttle type.
 	// Used for diagnosis: tells WHY the GPU throttled, not just whether it did.
-	ThermalThrottlePct  float64 `json:"thermal_throttle_pct"`  // HW+SW thermal slowdown
-	PowerCapThrottlePct float64 `json:"power_cap_throttle_pct"` // SW power cap
+	ThermalThrottlePct   float64 `json:"thermal_throttle_pct"`   // HW+SW thermal slowdown
+	PowerCapThrottlePct  float64 `json:"power_cap_throttle_pct"` // SW power cap
 	SyncBoostThrottlePct float64 `json:"sync_boost_throttle_pct,omitempty"`

 	// Temperature headroom: distance to the 100°C destruction threshold.
@@ -300,18 +301,22 @@ type NvidiaPowerBenchResult struct {
 	// PlatformMaxTDPW is the sum of per-GPU stable power limits found during the
 	// cumulative thermal ramp. Represents the actual sustained power budget of
 	// this server under full GPU load. Use for rack power planning.
-	PlatformMaxTDPW float64  `json:"platform_max_tdp_w"`
-	Findings        []string `json:"findings,omitempty"`
-	GPUs            []NvidiaPowerBenchGPU `json:"gpus"`
+	PlatformMaxTDPW float64 `json:"platform_max_tdp_w"`
+	// ServerPower captures IPMI server power delta (idle→loaded) measured in
+	// parallel with the thermal ramp. Use to compare GPU-reported TDP against
+	// actual wall-power draw as seen by the server's power supply.
+	ServerPower *BenchmarkServerPower `json:"server_power,omitempty"`
+	Findings    []string              `json:"findings,omitempty"`
+	GPUs        []NvidiaPowerBenchGPU `json:"gpus"`
 }

 type NvidiaPowerBenchGPU struct {
-	Index               int      `json:"index"`
-	Name                string   `json:"name,omitempty"`
-	BusID               string   `json:"bus_id,omitempty"`
-	DefaultPowerLimitW  float64  `json:"default_power_limit_w,omitempty"`
+	Index              int     `json:"index"`
+	Name               string  `json:"name,omitempty"`
+	BusID              string  `json:"bus_id,omitempty"`
+	DefaultPowerLimitW float64 `json:"default_power_limit_w,omitempty"`
 	// AppliedPowerLimitW is the stable limit found during single-card calibration.
-	AppliedPowerLimitW  float64  `json:"applied_power_limit_w,omitempty"`
+	AppliedPowerLimitW float64 `json:"applied_power_limit_w,omitempty"`
 	// StablePowerLimitW is the final fixed limit for this GPU after the
 	// cumulative thermal ramp. This is the limit at which the GPU operated
 	// stably with all other GPUs running simultaneously at their own limits.
@@ -326,13 +331,20 @@ type NvidiaPowerBenchGPU struct {
 	Notes               []string `json:"notes,omitempty"`
 	// CoolingWarning mirrors BenchmarkGPUResult.CoolingWarning for the power workflow.
 	CoolingWarning string `json:"cooling_warning,omitempty"`
+	// ServerLoadedW is the IPMI server power reading captured during this
+	// GPU's single-card calibration run. ServerDeltaW = ServerLoadedW − idle.
+	ServerLoadedW float64 `json:"server_loaded_w,omitempty"`
+	ServerDeltaW  float64 `json:"server_delta_w,omitempty"`
+	// Telemetry holds the aggregated stats from the final converged calibration
+	// attempt for this GPU (temperature, power, fan, clock percentiles).
+	Telemetry *BenchmarkTelemetrySummary `json:"telemetry,omitempty"`
 }

 type NvidiaPowerBenchStep struct {
-	StepIndex           int      `json:"step_index"`
-	GPUIndices          []int    `json:"gpu_indices"`
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
 	// NewGPUIndex is the GPU whose stable limit was searched in this step.
-	NewGPUIndex         int      `json:"new_gpu_index"`
+	NewGPUIndex int `json:"new_gpu_index"`
 	// NewGPUStableLimitW is the stable power limit found for the new GPU.
 	NewGPUStableLimitW  float64  `json:"new_gpu_stable_limit_w,omitempty"`
 	TotalObservedPowerW float64  `json:"total_observed_power_w,omitempty"`
@@ -340,20 +352,24 @@ type NvidiaPowerBenchStep struct {
 	Derated             bool     `json:"derated,omitempty"`
 	Status              string   `json:"status"`
 	Notes               []string `json:"notes,omitempty"`
+	// ServerLoadedW is the IPMI server power reading captured during this
+	// ramp step's calibration run. ServerDeltaW = ServerLoadedW − idle.
+	ServerLoadedW float64 `json:"server_loaded_w,omitempty"`
+	ServerDeltaW  float64 `json:"server_delta_w,omitempty"`
 }

 // NvidiaPerformanceRampStep holds per-step performance data for the
 // scalability ramp-up phase of the performance benchmark.
 type NvidiaPerformanceRampStep struct {
-	StepIndex          int      `json:"step_index"`
-	GPUIndices         []int    `json:"gpu_indices"`
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
 	// TotalSyntheticTOPS is the sum of per-GPU SyntheticScore (fp32-equivalent
 	// TOPS from dedicated single-precision phases) across all GPUs in this step.
-	TotalSyntheticTOPS float64  `json:"total_synthetic_tops"`
-	TotalMixedTOPS     float64  `json:"total_mixed_tops,omitempty"`
+	TotalSyntheticTOPS float64 `json:"total_synthetic_tops"`
+	TotalMixedTOPS     float64 `json:"total_mixed_tops,omitempty"`
 	// ScalabilityPct = TotalSyntheticTOPS / (k × best_single_gpu_tops) × 100.
 	// 100% = perfect linear scaling. < 100% = thermal/power/interconnect loss.
-	ScalabilityPct     float64  `json:"scalability_pct"`
-	Status             string   `json:"status"`
-	Notes              []string `json:"notes,omitempty"`
+	ScalabilityPct float64  `json:"scalability_pct"`
+	Status         string   `json:"status"`
+	Notes          []string `json:"notes,omitempty"`
 }
--- a/audit/internal/platform/gpu_metrics.go
+++ b/audit/internal/platform/gpu_metrics.go
@@ -27,6 +27,7 @@ type GPUMetricRow struct {
 	FanAvgRPM             float64 `json:"fan_avg_rpm,omitempty"`
 	FanDutyCyclePct       float64 `json:"fan_duty_cycle_pct,omitempty"`
 	FanDutyCycleAvailable bool    `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool    `json:"fan_duty_cycle_estimated,omitempty"`
 }

 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
@@ -147,14 +148,18 @@ 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("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")
+	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz,fan_avg_rpm,fan_duty_cycle_pct,fan_duty_cycle_available,fan_duty_cycle_estimated\n")
 	for _, r := range rows {
 		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)
+		dutyEstimated := 0
+		if r.FanDutyCycleEstimated {
+			dutyEstimated = 1
+		}
+		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%d,%d\n",
+			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz, r.FanAvgRPM, r.FanDutyCyclePct, dutyAvail, dutyEstimated)
 	}
 	return os.WriteFile(path, b.Bytes(), 0644)
 }
--- a/audit/internal/platform/install_to_ram.go
+++ b/audit/internal/platform/install_to_ram.go
@@ -140,26 +140,56 @@ func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) (ret
 	}

 	squashfsFiles, err := filepath.Glob("/run/live/medium/live/*.squashfs")
-	if err != nil || len(squashfsFiles) == 0 {
-		return fmt.Errorf("no squashfs files found in /run/live/medium/live/")
-	}
-
-	free := freeMemBytes()
-	var needed int64
-	for _, sf := range squashfsFiles {
-		fi, err2 := os.Stat(sf)
-		if err2 != nil {
-			return fmt.Errorf("stat %s: %v", sf, err2)
-		}
-		needed += fi.Size()
-	}
-	const headroom = 256 * 1024 * 1024
-	if free > 0 && needed+headroom > free {
-		return fmt.Errorf("insufficient RAM: need %s, available %s",
-			humanBytes(needed+headroom), humanBytes(free))
-	}
+	sourceAvailable := err == nil && len(squashfsFiles) > 0

 	dstDir := installToRAMDir
+
+	// If the source medium is unavailable, check whether a previous run already
+	// produced a complete copy in RAM. If so, skip the copy phase and proceed
+	// directly to the loop-rebind / bind-mount steps.
+	if !sourceAvailable {
+		copiedFiles, _ := filepath.Glob(filepath.Join(dstDir, "*.squashfs"))
+		if len(copiedFiles) > 0 {
+			log("Source medium not available, but a previous RAM copy was found — resuming from existing copy.")
+			// Proceed to rebind with the already-copied files.
+			for _, dst := range copiedFiles {
+				base := filepath.Base(dst)
+				// Re-associate the loop device that was originally backed by the
+				// source file (now gone); find it by the old source path pattern.
+				srcGuess := "/run/live/medium/live/" + base
+				loopDev, lerr := findLoopForFile(srcGuess)
+				if lerr != nil {
+					log(fmt.Sprintf("Loop device for %s not found (%v) — skipping re-association.", base, lerr))
+					continue
+				}
+				if rerr := reassociateLoopDevice(loopDev, dst); rerr != nil {
+					log(fmt.Sprintf("Warning: could not re-associate %s → %s: %v", loopDev, dst, rerr))
+				} else {
+					log(fmt.Sprintf("Loop device %s now backed by RAM copy.", loopDev))
+				}
+			}
+			goto bindMedium
+		}
+		return fmt.Errorf("no squashfs files found in /run/live/medium/live/ and no prior RAM copy in %s — reconnect the installation medium and retry", dstDir)
+	}
+
+	{
+		free := freeMemBytes()
+		var needed int64
+		for _, sf := range squashfsFiles {
+			fi, err2 := os.Stat(sf)
+			if err2 != nil {
+				return fmt.Errorf("stat %s: %v", sf, err2)
+			}
+			needed += fi.Size()
+		}
+		const headroom = 256 * 1024 * 1024
+		if free > 0 && needed+headroom > free {
+			return fmt.Errorf("insufficient RAM: need %s, available %s",
+				humanBytes(needed+headroom), humanBytes(free))
+		}
+	}
+
 	if state.CopyPresent {
 		log("Removing stale partial RAM copy before retry...")
 	}
@@ -199,6 +229,7 @@ func (s *System) RunInstallToRAM(ctx context.Context, logFunc func(string)) (ret
 		}
 	}

+bindMedium:
 	log("Copying remaining medium files...")
 	if err := cpDir(ctx, "/run/live/medium", dstDir, log); err != nil {
 		log(fmt.Sprintf("Warning: partial copy: %v", err))
--- a/audit/internal/platform/sat.go
+++ b/audit/internal/platform/sat.go
@@ -366,12 +366,14 @@ func (s *System) ResetNvidiaGPU(index int) (string, error) {
 	return string(raw), err
 }

-// RunNCCLTests runs nccl-tests all_reduce_perf across all NVIDIA GPUs.
+// RunNCCLTests runs nccl-tests all_reduce_perf across the selected NVIDIA GPUs.
 // Measures collective communication bandwidth over NVLink/PCIe.
-func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-	// detect GPU count
-	out, _ := exec.Command("nvidia-smi", "--query-gpu=index", "--format=csv,noheader").Output()
-	gpuCount := len(strings.Split(strings.TrimSpace(string(out)), "\n"))
+func (s *System) RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	gpuCount := len(selected)
 	if gpuCount < 1 {
 		gpuCount = 1
 	}
@@ -380,7 +382,7 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 		satJob{name: "02-all-reduce-perf.log", cmd: []string{
 			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
 			"-g", strconv.Itoa(gpuCount), "--iters", "20",
-		}},
+		}, env: nvidiaVisibleDevicesEnv(selected)},
 	), logFunc)
 }

--- a/audit/internal/platform/sat_fan_stress.go
+++ b/audit/internal/platform/sat_fan_stress.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"math"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -56,13 +57,37 @@ type cachedPowerReading struct {
 	UpdatedAt time.Time
 }

+type fanObservationState struct {
+	MaxRPM map[string]float64 `json:"max_rpm"`
+}
+
+type fanPeakCandidate struct {
+	FirstSeen time.Time
+	RPM       float64
+}
+
 var (
 	systemPowerCacheMu sync.Mutex
 	systemPowerCache   cachedPowerReading
+	fanObservationMu   sync.Mutex
+	fanObservation     fanObservationState
+	fanObservationInit bool
+	fanPeakCandidates  = make(map[string]fanPeakCandidate)
 )

 const systemPowerHoldTTL = 15 * time.Second

+var fanObservationStatePath = "/var/log/bee-sat/fan-observation.json"
+
+const fanObservationMinPeakHold = time.Second
+
+func normalizeObservedFanMaxRPM(rpm float64) float64 {
+	if rpm <= 0 {
+		return 0
+	}
+	return math.Ceil(rpm/1000.0) * 1000.0
+}
+
 // RunFanStressTest runs a two-phase GPU stress test while monitoring fan speeds,
 // temperatures, and power draw every second. Exports metrics.csv and fan-sensors.csv.
 // Designed to reproduce case-04 fan-speed lag and detect GPU thermal throttling.
@@ -310,11 +335,13 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	out, err := exec.Command("ipmitool", "sdr", "type", "Fan").Output()
 	if err == nil {
 		if fans := parseFanSpeeds(string(out)); len(fans) > 0 {
+			updateFanObservation(fans, time.Now())
 			return fans, nil
 		}
 	}
 	fans, sensorsErr := sampleFanSpeedsViaSensorsJSON()
 	if len(fans) > 0 {
+		updateFanObservation(fans, time.Now())
 		return fans, nil
 	}
 	if err != nil {
@@ -323,6 +350,119 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	return nil, sensorsErr
 }

+func loadFanObservationLocked() {
+	if fanObservationInit {
+		return
+	}
+	fanObservationInit = true
+	fanObservation.MaxRPM = make(map[string]float64)
+	raw, err := os.ReadFile(fanObservationStatePath)
+	if err != nil || len(raw) == 0 {
+		return
+	}
+	var persisted fanObservationState
+	if json.Unmarshal(raw, &persisted) != nil {
+		return
+	}
+	for name, rpm := range persisted.MaxRPM {
+		name = strings.TrimSpace(name)
+		if name == "" || rpm <= 0 {
+			continue
+		}
+		fanObservation.MaxRPM[name] = rpm
+	}
+}
+
+func saveFanObservationLocked() {
+	if len(fanObservation.MaxRPM) == 0 {
+		return
+	}
+	dir := filepath.Dir(fanObservationStatePath)
+	if dir == "" || dir == "." {
+		dir = "/var/log/bee-sat"
+	}
+	if err := os.MkdirAll(dir, 0755); err != nil {
+		return
+	}
+	raw, err := json.MarshalIndent(fanObservation, "", "  ")
+	if err != nil {
+		return
+	}
+	_ = os.WriteFile(fanObservationStatePath, raw, 0644)
+}
+
+func updateFanObservation(fans []FanReading, now time.Time) {
+	if len(fans) == 0 {
+		return
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	changed := false
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		currentMax := fanObservation.MaxRPM[name]
+		if fan.RPM <= currentMax {
+			delete(fanPeakCandidates, name)
+			continue
+		}
+		if cand, ok := fanPeakCandidates[name]; ok {
+			if now.Sub(cand.FirstSeen) >= fanObservationMinPeakHold {
+				newMax := math.Max(cand.RPM, fan.RPM)
+				if newMax > currentMax {
+					fanObservation.MaxRPM[name] = normalizeObservedFanMaxRPM(newMax)
+					changed = true
+				}
+				delete(fanPeakCandidates, name)
+				continue
+			}
+			if fan.RPM > cand.RPM {
+				fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: cand.FirstSeen, RPM: fan.RPM}
+			}
+			continue
+		}
+		fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: now, RPM: fan.RPM}
+	}
+	if changed {
+		saveFanObservationLocked()
+	}
+}
+
+func estimateFanDutyCyclePctFromObservation(fans []FanReading) (float64, bool) {
+	if len(fans) == 0 {
+		return 0, false
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	var samples []float64
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		maxRPM := fanObservation.MaxRPM[name]
+		if maxRPM <= 0 {
+			continue
+		}
+		pct := fan.RPM / maxRPM * 100.0
+		if pct > 100 {
+			pct = 100
+		}
+		if pct < 0 {
+			pct = 0
+		}
+		samples = append(samples, pct)
+	}
+	if len(samples) == 0 {
+		return 0, false
+	}
+	return benchmarkMean(samples), true
+}
+
 // parseFanSpeeds parses "ipmitool sdr type Fan" output.
 // Handles two formats:
 //
@@ -428,12 +568,27 @@ func sampleFanSpeedsViaSensorsJSON() ([]FanReading, error) {

 // sampleFanDutyCyclePct reads fan PWM/duty-cycle controls from lm-sensors.
 // Returns the average duty cycle across all exposed PWM controls.
-func sampleFanDutyCyclePct() (float64, bool) {
+func sampleFanDutyCyclePct() (float64, bool, bool) {
 	out, err := exec.Command("sensors", "-j").Output()
 	if err != nil || len(out) == 0 {
-		return 0, false
+		fans, fanErr := sampleFanSpeeds()
+		if fanErr != nil {
+			return 0, false, false
+		}
+		return sampleFanDutyCyclePctFromFans(fans)
 	}
-	return parseFanDutyCyclePctSensorsJSON(out)
+	pct, ok := parseFanDutyCyclePctSensorsJSON(out)
+	return pct, ok, false
+}
+
+func sampleFanDutyCyclePctFromFans(fans []FanReading) (float64, bool, bool) {
+	if len(fans) == 0 {
+		return 0, false, false
+	}
+	if pct, ok := estimateFanDutyCyclePctFromObservation(fans); ok {
+		return pct, true, true
+	}
+	return 0, false, false
 }

 func parseFanDutyCyclePctSensorsJSON(raw []byte) (float64, bool) {
--- a/audit/internal/platform/sat_fan_stress_test.go
+++ b/audit/internal/platform/sat_fan_stress_test.go
@@ -1,6 +1,7 @@
 package platform

 import (
+	"path/filepath"
 	"testing"
 	"time"
 )
@@ -50,6 +51,53 @@ func TestParseFanDutyCyclePctSensorsJSON(t *testing.T) {
 	}
 }

+func TestEstimateFanDutyCyclePctFromObservation(t *testing.T) {
+	t.Parallel()
+
+	oldPath := fanObservationStatePath
+	oldState := fanObservation
+	oldInit := fanObservationInit
+	oldCandidates := fanPeakCandidates
+	fanObservationStatePath = filepath.Join(t.TempDir(), "fan-observation.json")
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	t.Cleanup(func() {
+		fanObservationStatePath = oldPath
+		fanObservation = oldState
+		fanObservationInit = oldInit
+		fanPeakCandidates = oldCandidates
+	})
+
+	start := time.Unix(100, 0)
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5000}}, start)
+	if _, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2500}}); ok {
+		t.Fatalf("single-sample spike should not establish observed max")
+	}
+
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5200}}, start.Add(500*time.Millisecond))
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5100}}, start.Add(1500*time.Millisecond))
+
+	got, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected estimated duty cycle from persisted observed max")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("got=%v want ~43.3", got)
+	}
+
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	got, ok = estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected persisted observed max to be reloaded from disk")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("reloaded got=%v want ~43.3", got)
+	}
+}
+
 func TestParseDCMIPowerReading(t *testing.T) {
 	raw := `
 Instantaneous power reading:                   512 Watts
--- a/audit/internal/platform/sat_test.go
+++ b/audit/internal/platform/sat_test.go
@@ -321,6 +321,19 @@ func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
 	}
 }

+func TestNvidiaDCGMNamedDiagCommandSkipsDurationForNVBandwidth(t *testing.T) {
+	cmd := nvidiaDCGMNamedDiagCommand("nvbandwidth", 0, []int{2, 0})
+	want := []string{"dcgmi", "diag", "-r", "nvbandwidth", "-i", "2,0"}
+	if len(cmd) != len(want) {
+		t.Fatalf("cmd len=%d want %d (%v)", len(cmd), len(want), cmd)
+	}
+	for i := range want {
+		if cmd[i] != want[i] {
+			t.Fatalf("cmd[%d]=%q want %q", i, cmd[i], want[i])
+		}
+	}
+}
+
 func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
 	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
 	if len(env) != 2 {
--- a/audit/internal/webui/pages.go
+++ b/audit/internal/webui/pages.go
@@ -1481,7 +1481,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Verifies NVLink/NVSwitch fabric bandwidth using NCCL all_reduce_perf across all selected GPUs. Pass/fail based on achieved bandwidth vs. theoretical.`,
 			`<code>all_reduce_perf</code> (NCCL tests)`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs across all selected GPUs simultaneously (requires ≥2).<p id="sat-ni-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			`Runs in Validate and Stress. Uses all selected GPUs simultaneously (requires ≥2) and is kept short so it fits the Validate flow.`,
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-bandwidth">` +
@@ -1489,7 +1489,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Validates GPU memory copy and peer-to-peer bandwidth paths using NVBandwidth.`,
 			`<code>nvbandwidth</code>`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs across all selected GPUs simultaneously.<p id="sat-nb-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			`Runs in Validate and Stress across all selected GPUs simultaneously. Intended to stay short enough for Validate.`,
 		)) +
 		`</div>` +
 		`</div>
@@ -1527,8 +1527,6 @@ function satModeChanged() {
    {card: 'sat-card-nvidia-targeted-stress', hint: 'sat-ts-mode-hint'},
    {card: 'sat-card-nvidia-targeted-power',  hint: 'sat-tp-mode-hint'},
    {card: 'sat-card-nvidia-pulse',           hint: 'sat-pt-mode-hint'},
-    {card: 'sat-card-nvidia-interconnect',    hint: 'sat-ni-mode-hint'},
-    {card: 'sat-card-nvidia-bandwidth',       hint: 'sat-nb-mode-hint'},
  ].forEach(function(item) {
    const card = document.getElementById(item.card);
    if (card) {
@@ -1776,7 +1774,7 @@ function runAllSAT() {
  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'];
+  const stressOnlyTargets = ['nvidia-targeted-stress', 'nvidia-targeted-power', 'nvidia-pulse'];
  const baseTargets = ['nvidia','nvidia-targeted-stress','nvidia-targeted-power','nvidia-pulse','nvidia-interconnect','nvidia-bandwidth','memory','storage','cpu'].concat(selectedAMDValidateTargets());
  const activeTargets = baseTargets.filter(target => {
    if (stressOnlyTargets.indexOf(target) >= 0 && !satStressMode()) return false;
@@ -2016,9 +2014,11 @@ func renderSATCard(id, label, runAction, headerActions, body string) string {
 // ── Benchmark ─────────────────────────────────────────────────────────────────

 type benchmarkHistoryRun struct {
-	generatedAt time.Time
-	displayTime string
-	gpuScores   map[int]float64 // GPU index → composite score
+	generatedAt  time.Time
+	displayTime  string
+	gpuScores    map[int]float64 // GPU index → composite score
+	gpuStatuses  map[int]string  // GPU index → status ("OK", "WARNING", "FAILED", …)
+	overallStatus string
 }

 func renderBenchmark(opts HandlerOptions) string {
@@ -2082,7 +2082,7 @@ func renderBenchmark(opts HandlerOptions) string {
  </div>
 </div>

-`+`<div id="benchmark-results-section">`+renderBenchmarkResultsCard(opts.ExportDir)+`</div>`+`
+` + `<div id="benchmark-results-section">` + renderBenchmarkResultsCard(opts.ExportDir) + `</div>` + `

 <div id="benchmark-output" style="display:none;margin-top:16px" class="card">
  <div class="card-head">Benchmark Output <span id="benchmark-title"></span></div>
@@ -2326,7 +2326,7 @@ 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>Run</th><th>Time</th>`)
+	b.WriteString(`<table><thead><tr><th>Run</th><th>Time</th><th>Status</th>`)
 	for i := 0; i <= maxGPUIndex; i++ {
 		b.WriteString(`<th>GPU ` + strconv.Itoa(i) + `</th>`)
 	}
@@ -2335,13 +2335,36 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 		b.WriteString(`<tr>`)
 		b.WriteString(`<td>#` + strconv.Itoa(i+1) + `</td>`)
 		b.WriteString(`<td>` + html.EscapeString(run.displayTime) + `</td>`)
+		overallColor := "var(--ok)"
+		overallLabel := run.overallStatus
+		if overallLabel == "" {
+			overallLabel = "OK"
+		}
+		if overallLabel == "FAILED" {
+			overallColor = "var(--crit-fg,#9f3a38)"
+		} else if overallLabel != "OK" {
+			overallColor = "var(--warn)"
+		}
+		b.WriteString(`<td style="color:` + overallColor + `;font-weight:600">` + html.EscapeString(overallLabel) + `</td>`)
 		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", score) + `</td>`)
+			gpuStatus := run.gpuStatuses[idx]
+			scoreColor := ""
+			switch gpuStatus {
+			case "FAILED":
+				scoreColor = ` style="color:var(--crit-fg,#9f3a38);font-weight:600"`
+			case "WARNING", "PARTIAL":
+				scoreColor = ` style="color:var(--warn);font-weight:600"`
+			case "", "OK":
+				// no override
+			default:
+				scoreColor = ` style="color:var(--warn);font-weight:600"`
+			}
+			b.WriteString(`<td` + scoreColor + `>` + fmt.Sprintf("%.2f", score) + `</td>`)
 		}
 		b.WriteString(`</tr>`)
 	}
@@ -2375,12 +2398,15 @@ func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun)
 			continue
 		}
 		run := benchmarkHistoryRun{
-			generatedAt: result.GeneratedAt,
-			displayTime: result.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
-			gpuScores:   make(map[int]float64),
+			generatedAt:   result.GeneratedAt,
+			displayTime:   result.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
+			gpuScores:     make(map[int]float64),
+			gpuStatuses:   make(map[int]string),
+			overallStatus: result.OverallStatus,
 		}
 		for _, gpu := range result.GPUs {
 			run.gpuScores[gpu.Index] = gpu.Scores.CompositeScore
+			run.gpuStatuses[gpu.Index] = gpu.Status
 			if gpu.Index > maxGPUIndex {
 				maxGPUIndex = gpu.Index
 			}
@@ -2449,31 +2475,45 @@ func renderPowerBenchmarkResultsCard(exportDir string) string {

 	if len(latest.GPUs) > 0 {
 		b.WriteString(`<div style="overflow-x:auto"><table><thead><tr>`)
-		b.WriteString(`<th>GPU</th><th>Model</th><th>Nominal W</th><th>Achieved W</th><th>P95 Observed W</th><th>Status</th>`)
+		b.WriteString(`<th>GPU</th><th>Model</th><th>Nominal W</th><th>Single-card W</th><th>Multi-GPU W</th><th>P95 Observed W</th><th>Status</th>`)
 		b.WriteString(`</tr></thead><tbody>`)
 		for _, gpu := range latest.GPUs {
-			derated := gpu.Derated || (gpu.DefaultPowerLimitW > 0 && gpu.AppliedPowerLimitW < gpu.DefaultPowerLimitW-1)
+			// finalLimitW is the definitive TDP: multi-GPU stable limit from the ramp,
+			// falling back to single-card applied limit if the ramp hasn't run.
+			finalLimitW := gpu.StablePowerLimitW
+			if finalLimitW <= 0 {
+				finalLimitW = gpu.AppliedPowerLimitW
+			}
+			// Derate is relative to nominal (DefaultPowerLimitW), using the final limit.
+			derated := gpu.Derated ||
+				(gpu.DefaultPowerLimitW > 0 && finalLimitW > 0 && finalLimitW < gpu.DefaultPowerLimitW-1)
 			rowStyle := ""
-			achievedStyle := ""
+			finalStyle := ""
 			if derated {
 				rowStyle = ` style="background:rgba(255,180,0,0.08)"`
-				achievedStyle = ` style="color:#e6a000;font-weight:600"`
+				finalStyle = ` style="color:#e6a000;font-weight:600"`
 			}
 			statusLabel := gpu.Status
 			if statusLabel == "" {
 				statusLabel = "OK"
 			}
 			statusColor := "var(--ok)"
-			if statusLabel != "OK" {
+			if statusLabel == "FAILED" {
+				statusColor = "var(--crit-fg,#9f3a38)"
+			} else if statusLabel != "OK" {
 				statusColor = "var(--warn)"
 			}
 			nominalStr := "-"
 			if gpu.DefaultPowerLimitW > 0 {
 				nominalStr = fmt.Sprintf("%.0f", gpu.DefaultPowerLimitW)
 			}
-			achievedStr := "-"
+			singleStr := "-"
 			if gpu.AppliedPowerLimitW > 0 {
-				achievedStr = fmt.Sprintf("%.0f", gpu.AppliedPowerLimitW)
+				singleStr = fmt.Sprintf("%.0f", gpu.AppliedPowerLimitW)
+			}
+			multiStr := "-"
+			if gpu.StablePowerLimitW > 0 {
+				multiStr = fmt.Sprintf("%.0f", gpu.StablePowerLimitW)
 			}
 			p95Str := "-"
 			if gpu.MaxObservedPowerW > 0 {
@@ -2483,7 +2523,8 @@ func renderPowerBenchmarkResultsCard(exportDir string) string {
 			b.WriteString(`<td>` + strconv.Itoa(gpu.Index) + `</td>`)
 			b.WriteString(`<td>` + html.EscapeString(gpu.Name) + `</td>`)
 			b.WriteString(`<td>` + nominalStr + `</td>`)
-			b.WriteString(`<td` + achievedStyle + `>` + achievedStr + `</td>`)
+			b.WriteString(`<td>` + singleStr + `</td>`)
+			b.WriteString(`<td` + finalStyle + `>` + multiStr + `</td>`)
 			b.WriteString(`<td>` + p95Str + `</td>`)
 			b.WriteString(`<td style="color:` + statusColor + `;font-weight:600">` + html.EscapeString(statusLabel) + `</td>`)
 			b.WriteString(`</tr>`)
@@ -2517,7 +2558,7 @@ func renderPowerBenchmarkResultsCard(exportDir string) string {

 func renderBurn() string {
 	return `<div class="alert alert-warn" style="margin-bottom:16px"><strong>&#9888; Warning:</strong> Stress tests on this page run hardware at high load. Repeated or prolonged use may reduce hardware lifespan. Use only when necessary.</div>
-<div class="alert alert-info" style="margin-bottom:16px"><strong>Scope:</strong> DCGM diagnostics (` + "targeted_stress, targeted_power, pulse_test" + `), NCCL, NVBandwidth, and LINPACK remain in <a href="/validate">Validate → Stress mode</a>. Burn exposes sustained GPU compute load recipes.</div>
+<div class="alert alert-info" style="margin-bottom:16px"><strong>Scope:</strong> Burn exposes sustained GPU compute load recipes. DCGM diagnostics (` + "targeted_stress, targeted_power, pulse_test" + `) and LINPACK remain in <a href="/validate">Validate → Stress mode</a>; NCCL and NVBandwidth are available directly from <a href="/validate">Validate</a>.</div>
 <p style="color:var(--muted);font-size:13px;margin-bottom:16px">Tasks continue in the background — view progress in <a href="/tasks">Tasks</a>.</p>

 <div class="card" style="margin-bottom:16px">
--- a/audit/internal/webui/server_test.go
+++ b/audit/internal/webui/server_test.go
@@ -744,6 +744,26 @@ func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
 	}
 }

+func TestValidatePageRendersNvidiaFabricCardsInValidateMode(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/validate", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`NVIDIA Interconnect (NCCL)`,
+		`Runs in Validate and Stress.`,
+		`NVIDIA Bandwidth (NVBandwidth)`,
+		`Intended to stay short enough for Validate.`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("validate page missing %q: %s", needle, body)
+		}
+	}
+}
+
 func TestBurnPageRendersGoalBasedNVIDIACards(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
--- a/audit/internal/webui/tasks.go
+++ b/audit/internal/webui/tasks.go
@@ -613,8 +613,9 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 	}
 	a := q.opts.App

+	recovered := len(j.lines) > 0
 	j.append(fmt.Sprintf("Starting %s...", t.Name))
-	if len(j.lines) > 0 {
+	if recovered {
 		j.append(fmt.Sprintf("Recovered after bee-web restart at %s", time.Now().UTC().Format(time.RFC3339)))
 	}

@@ -736,15 +737,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		dur := t.params.Duration
-		if t.params.BurnProfile != "" && dur <= 0 {
-			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
-		}
-		archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
-			DurationSec: dur,
-			Loader:      platform.NvidiaStressLoaderNCCL,
-			GPUIndices:  t.params.GPUIndices,
-		}, j.append)
+		archive, err = a.RunNCCLTests(ctx, "", t.params.GPUIndices, j.append)
 	case "nvidia-stress":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
--- a/iso/builder/bee-gpu-stress.c
+++ b/iso/builder/bee-gpu-stress.c
@@ -713,6 +713,19 @@ static const struct profile_desc k_profiles[] = {

 #define PROFILE_COUNT ((int)(sizeof(k_profiles) / sizeof(k_profiles[0])))

+static int profile_allowed_for_run(const struct profile_desc *desc, int cc, const char *precision_filter) {
+    if (!(desc->enabled && cc >= desc->min_cc)) {
+        return 0;
+    }
+    if (precision_filter != NULL) {
+        return strcmp(desc->block_label, precision_filter) == 0;
+    }
+    /* Mixed/all phases intentionally exclude fp64/fp4 for now: both paths are
+     * unstable on the current benchmark fleet and can abort the whole mixed
+     * pass after earlier phases already collected useful telemetry. */
+    return strcmp(desc->block_label, "fp64") != 0 && strcmp(desc->block_label, "fp4") != 0;
+}
+
 static int load_cublaslt(struct cublaslt_api *api) {
    memset(api, 0, sizeof(*api));
    api->lib = dlopen("libcublasLt.so.13", RTLD_NOW | RTLD_LOCAL);
@@ -1222,8 +1235,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,

    /* 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 &&
-            (precision_filter == NULL || strcmp(k_profiles[i].block_label, precision_filter) == 0)) {
+        if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
            planned++;
        }
    }
@@ -1240,7 +1252,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     * profiles matching precision_filter. */
    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) {
+        if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
            planned_total++;
        }
    }
@@ -1310,10 +1322,10 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                          desc->min_cc);
            continue;
        }
-        if (precision_filter != NULL && strcmp(desc->block_label, precision_filter) != 0) {
+        if (!profile_allowed_for_run(desc, cc, precision_filter)) {
            append_detail(report->details,
                          sizeof(report->details),
-                          "%s=SKIPPED precision_filter\n",
+                          "%s=SKIPPED benchmark_disabled\n",
                          desc->name);
            continue;
        }
--- a/iso/builder/config/bootloaders/grub-pc/grub.cfg
+++ b/iso/builder/config/bootloaders/grub-pc/grub.cfg
@@ -16,6 +16,11 @@ menuentry "EASY-BEE" {
 }

 submenu "EASY-BEE (advanced options) -->" {
+    menuentry "EASY-BEE — load to RAM (toram)" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ toram nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
+
    menuentry "EASY-BEE — GSP=off" {
        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable pcie_aspm=off intel_idle.max_cstate=1 processor.max_cstate=1 nowatchdog nosoftlockup
        initrd  @INITRD_LIVE@
--- a/iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot
+++ b/iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot
@@ -63,8 +63,10 @@ chmod +x /usr/local/bin/bee-sshsetup   2>/dev/null || true
 chmod +x /usr/local/bin/bee-smoketest  2>/dev/null || true
 chmod +x /usr/local/bin/bee            2>/dev/null || true
 chmod +x /usr/local/bin/bee-log-run    2>/dev/null || true
-chmod +x /usr/local/bin/bee-selfheal      2>/dev/null || true
-chmod +x /usr/local/bin/bee-boot-status  2>/dev/null || true
+chmod +x /usr/local/bin/bee-selfheal        2>/dev/null || true
+chmod +x /usr/local/bin/bee-boot-status    2>/dev/null || true
+chmod +x /usr/local/bin/bee-install        2>/dev/null || true
+chmod +x /usr/local/bin/bee-remount-medium 2>/dev/null || true
 if [ "$GPU_VENDOR" = "nvidia" ]; then
    chmod +x /usr/local/bin/bee-nvidia-load 2>/dev/null || true
    chmod +x /usr/local/bin/bee-gpu-burn 2>/dev/null || true
--- a/iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot
+++ b/iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot
@@ -1,117 +0,0 @@
-#!/bin/sh
-# 9001-wallpaper.hook.chroot — generate /usr/share/bee/wallpaper.png inside chroot
-set -e
-echo "=== generating bee wallpaper ==="
-mkdir -p /usr/share/bee
-
-python3 - <<'PYEOF'
-from PIL import Image, ImageDraw, ImageFont, ImageFilter
-import os
-
-W, H = 1920, 1080
-
-ASCII_ART = [
-    "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗",
-    "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝",
-    "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗",
-    "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝",
-    "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗",
-    "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝",
-]
-SUBTITLE = "  Hardware Audit LiveCD"
-
-FG = (0xF6, 0xD0, 0x47)
-FG_DIM = (0xD4, 0xA9, 0x1C)
-SHADOW = (0x5E, 0x47, 0x05)
-SUB = (0x96, 0x7A, 0x17)
-BG = (0x05, 0x05, 0x05)
-
-MONO_FONT_CANDIDATES = [
-    '/usr/share/fonts/truetype/dejavu/DejaVuSansMono-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation/LiberationMono-Bold.ttf',
-    '/usr/share/fonts/truetype/freefont/FreeMonoBold.ttf',
-]
-SUB_FONT_CANDIDATES = [
-    '/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation2/LiberationSans-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf',
-    '/usr/share/fonts/truetype/freefont/FreeSansBold.ttf',
-]
-
-
-def load_font(candidates, size):
-    for path in candidates:
-        if os.path.exists(path):
-            return ImageFont.truetype(path, size)
-    return ImageFont.load_default()
-
-
-def mono_metrics(font):
-    probe = Image.new('L', (W, H), 0)
-    draw = ImageDraw.Draw(probe)
-    char_w = int(round(draw.textlength("M", font=font)))
-    bb = draw.textbbox((0, 0), "Mg", font=font)
-    char_h = bb[3] - bb[1]
-    return char_w, char_h
-
-
-def render_ascii_mask(font, lines, char_w, char_h, line_gap):
-    width = max(len(line) for line in lines) * char_w
-    height = len(lines) * char_h + line_gap * (len(lines) - 1)
-    mask = Image.new('L', (width, height), 0)
-    draw = ImageDraw.Draw(mask)
-    for row, line in enumerate(lines):
-        y = row * (char_h + line_gap)
-        for col, ch in enumerate(line):
-            if ch == ' ':
-                continue
-            x = col * char_w
-            draw.text((x, y), ch, font=font, fill=255)
-    return mask
-
-
-img = Image.new('RGB', (W, H), BG)
-draw = ImageDraw.Draw(img)
-
-# Soft amber glow under the logo without depending on font rendering.
-glow = Image.new('RGBA', (W, H), (0, 0, 0, 0))
-glow_draw = ImageDraw.Draw(glow)
-glow_draw.ellipse((360, 250, 1560, 840), fill=(180, 120, 10, 56))
-glow_draw.ellipse((520, 340, 1400, 760), fill=(255, 190, 40, 36))
-glow = glow.filter(ImageFilter.GaussianBlur(60))
-img = Image.alpha_composite(img.convert('RGBA'), glow)
-
-TARGET_LOGO_W = 400
-max_chars = max(len(line) for line in ASCII_ART)
-_probe_font = load_font(MONO_FONT_CANDIDATES, 64)
-_probe_cw, _ = mono_metrics(_probe_font)
-font_size_logo = max(6, int(64 * TARGET_LOGO_W / (_probe_cw * max_chars)))
-font_logo = load_font(MONO_FONT_CANDIDATES, font_size_logo)
-char_w, char_h = mono_metrics(font_logo)
-logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 2)
-logo_w, logo_h = logo_mask.size
-logo_x = (W - logo_w) // 2
-logo_y = 380
-
-sh_off = max(1, font_size_logo // 6)
-shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(1))
-img.paste(SHADOW, (logo_x + sh_off * 2, logo_y + sh_off * 2), shadow_mask)
-img.paste(FG_DIM, (logo_x + sh_off, logo_y + sh_off), logo_mask)
-img.paste(FG, (logo_x, logo_y), logo_mask)
-
-font_sub = load_font(SUB_FONT_CANDIDATES, 30)
-sub_bb = draw.textbbox((0, 0), SUBTITLE, font=font_sub)
-sub_x = (W - (sub_bb[2] - sub_bb[0])) // 2
-sub_y = logo_y + logo_h + 48
-draw = ImageDraw.Draw(img)
-draw.text((sub_x + 2, sub_y + 2), SUBTITLE, font=font_sub, fill=(35, 28, 6))
-draw.text((sub_x, sub_y), SUBTITLE, font=font_sub, fill=SUB)
-
-img = img.convert('RGB')
-
-img.save('/usr/share/bee/wallpaper.png', optimize=True)
-print('wallpaper written: /usr/share/bee/wallpaper.png')
-PYEOF
-
-echo "=== wallpaper done ==="
--- a/iso/builder/config/hooks/normal/9011-toram-rsync.hook.chroot
+++ b/iso/builder/config/hooks/normal/9011-toram-rsync.hook.chroot
@@ -0,0 +1,46 @@
+#!/bin/sh
+# 9011-toram-rsync.hook.chroot
+#
+# Adds rsync to the initramfs so that live-boot's toram code takes the
+# rsync --progress path instead of the silent "cp -a" fallback.
+#
+# live-boot's 9990-toram-todisk.sh already contains:
+#   if [ -x /bin/rsync ]; then
+#       rsync -a --progress ... 1>/dev/console
+#   else
+#       cp -a ...   # no output
+#   fi
+#
+# We install an initramfs-tools hook that calls copy_exec /usr/bin/rsync,
+# which copies the binary + all shared-library dependencies into the initrd.
+
+set -e
+
+HOOK_DIR="/etc/initramfs-tools/hooks"
+HOOK="${HOOK_DIR}/bee-rsync"
+
+mkdir -p "${HOOK_DIR}"
+
+cat > "${HOOK}" << 'EOF'
+#!/bin/sh
+# initramfs hook: include rsync for live-boot toram progress output
+PREREQ=""
+prereqs() { echo "$PREREQ"; }
+case "$1" in prereqs) prereqs; exit 0 ;; esac
+
+. /usr/share/initramfs-tools/hook-functions
+
+if [ -x /usr/bin/rsync ]; then
+    copy_exec /usr/bin/rsync /bin
+fi
+EOF
+
+chmod +x "${HOOK}"
+
+echo "9011-toram-rsync: installed initramfs hook at ${HOOK}"
+
+# Rebuild initramfs so the hook takes effect in the ISO's initrd.img
+KVER=$(ls /lib/modules | sort -V | tail -1)
+echo "9011-toram-rsync: rebuilding initramfs for kernel ${KVER}"
+update-initramfs -u -k "${KVER}"
+echo "9011-toram-rsync: done"
--- a/iso/builder/config/package-lists/bee.list.chroot
+++ b/iso/builder/config/package-lists/bee.list.chroot
@@ -3,6 +3,7 @@ dmidecode
 smartmontools
 nvme-cli
 pciutils
+rsync
 ipmitool
 util-linux
 e2fsprogs
--- a/iso/overlay/usr/local/bin/bee-install
+++ b/iso/overlay/usr/local/bin/bee-install
@@ -65,6 +65,9 @@ done
 SQUASHFS="/run/live/medium/live/filesystem.squashfs"
 if [ ! -f "$SQUASHFS" ]; then
    echo "ERROR: squashfs not found at $SQUASHFS" >&2
+    echo "  The live medium may have been disconnected." >&2
+    echo "  Reconnect the disc and run:  bee-remount-medium --wait" >&2
+    echo "  Then re-run bee-install." >&2
    exit 1
 fi

@@ -162,10 +165,59 @@ log "  Mounted."
 log "--- Step 5/7: Unpacking filesystem (this takes 10-20 minutes) ---"
 log "  Source: $SQUASHFS"
 log "  Target: $MOUNT_ROOT"
-unsquashfs -f -d "$MOUNT_ROOT" "$SQUASHFS" 2>&1 | \
-    grep -E '^\[|^inod|^created|^extract' | \
-    while read -r line; do log "  $line"; done || true
-log "  Unpack complete."
+
+# unsquashfs does not support resume, so retry the entire unpack step if the
+# source medium disappears mid-copy (e.g. CD physically disconnected).
+UNPACK_ATTEMPTS=0
+UNPACK_MAX=5
+while true; do
+    UNPACK_ATTEMPTS=$(( UNPACK_ATTEMPTS + 1 ))
+    if [ "$UNPACK_ATTEMPTS" -gt "$UNPACK_MAX" ]; then
+        die "Unpack failed $UNPACK_MAX times — giving up. Check the disc and logs."
+    fi
+    [ "$UNPACK_ATTEMPTS" -gt 1 ] && log "  Retry attempt $UNPACK_ATTEMPTS / $UNPACK_MAX ..."
+
+    # Re-check squashfs is reachable before each attempt
+    if [ ! -f "$SQUASHFS" ]; then
+        log "  SOURCE LOST: $SQUASHFS not found."
+        log "  Reconnect the disc and run 'bee-remount-medium --wait' in another terminal,"
+        log "  then press Enter here to retry."
+        read -r _
+        continue
+    fi
+
+    # wipe partial unpack so unsquashfs starts clean
+    if [ "$UNPACK_ATTEMPTS" -gt 1 ]; then
+        log "  Cleaning partial unpack from $MOUNT_ROOT ..."
+        # keep the mount point itself but remove its contents
+        find "$MOUNT_ROOT" -mindepth 1 -maxdepth 1 -exec rm -rf {} + 2>/dev/null || true
+    fi
+
+    UNPACK_OK=0
+    unsquashfs -f -d "$MOUNT_ROOT" "$SQUASHFS" 2>&1 | \
+        grep -E '^\[|^inod|^created|^extract|^ERROR|failed' | \
+        while IFS= read -r line; do log "  $line"; done || UNPACK_OK=$?
+
+    # Check squashfs is still reachable (gone = disc pulled during copy)
+    if [ ! -f "$SQUASHFS" ]; then
+        log "  WARNING: source medium lost during unpack — will retry after remount."
+        log "  Run 'bee-remount-medium --wait' in another terminal, then press Enter."
+        read -r _
+        continue
+    fi
+
+    # Verify the unpack produced a usable root (presence of /etc is a basic check)
+    if [ -d "${MOUNT_ROOT}/etc" ]; then
+        log "  Unpack complete."
+        break
+    else
+        log "  WARNING: unpack produced no /etc — squashfs may be corrupt or incomplete."
+        if [ "$UNPACK_ATTEMPTS" -lt "$UNPACK_MAX" ]; then
+            log "  Retrying in 5 s ..."
+            sleep 5
+        fi
+    fi
+done

 # ------------------------------------------------------------------
 log "--- Step 6/7: Configuring installed system ---"
--- a/iso/overlay/usr/local/bin/bee-remount-medium
+++ b/iso/overlay/usr/local/bin/bee-remount-medium
@@ -0,0 +1,100 @@
+#!/bin/bash
+# bee-remount-medium — find and remount the live ISO medium to /run/live/medium
+#
+# Run this after reconnecting the ISO source disc (USB/CD) if the live medium
+# was lost and /run/live/medium/live/filesystem.squashfs is missing.
+#
+# Usage: bee-remount-medium [--wait]
+#   --wait  keep retrying every 5 seconds until the medium is found (useful
+#           while physically reconnecting the device)
+
+set -euo pipefail
+
+MEDIUM_DIR="/run/live/medium"
+SQUASHFS_REL="live/filesystem.squashfs"
+WAIT_MODE=0
+
+for arg in "$@"; do
+    case "$arg" in
+        --wait|-w) WAIT_MODE=1 ;;
+        --help|-h)
+            echo "Usage: bee-remount-medium [--wait]"
+            echo "  Finds and remounts the live ISO medium to $MEDIUM_DIR"
+            echo "  --wait  retry every 5 s until a medium with squashfs is found"
+            exit 0 ;;
+    esac
+done
+
+log() { echo "[$(date +%H:%M:%S)] $*"; }
+die() { log "ERROR: $*" >&2; exit 1; }
+
+# Return all candidate block devices (optical + removable USB mass storage)
+find_candidates() {
+    # CD/DVD drives
+    for dev in /dev/sr* /dev/scd*; do
+        [ -b "$dev" ] && echo "$dev"
+    done
+    # USB/removable disks and partitions
+    for dev in /dev/sd* /dev/vd*; do
+        [ -b "$dev" ] || continue
+        # Only whole disks or partitions — skip the same device we are running from
+        local removable
+        local base
+        base=$(basename "$dev")
+        removable=$(cat "/sys/block/${base%%[0-9]*}/removable" 2>/dev/null || echo 0)
+        [ "$removable" = "1" ] && echo "$dev"
+    done
+}
+
+# Try to mount $1 to $MEDIUM_DIR and check for squashfs
+try_mount() {
+    local dev="$1"
+    local tmpdir
+    tmpdir=$(mktemp -d /tmp/bee-probe-XXXXXX)
+    if mount -o ro "$dev" "$tmpdir" 2>/dev/null; then
+        if [ -f "${tmpdir}/${SQUASHFS_REL}" ]; then
+            # Unmount probe mount and mount properly onto live path
+            umount "$tmpdir" 2>/dev/null || true
+            rmdir "$tmpdir"  2>/dev/null || true
+            # Unmount whatever is currently on MEDIUM_DIR (may be empty/stale)
+            umount "$MEDIUM_DIR" 2>/dev/null || true
+            mkdir -p "$MEDIUM_DIR"
+            if mount -o ro "$dev" "$MEDIUM_DIR"; then
+                log "Mounted $dev on $MEDIUM_DIR"
+                return 0
+            else
+                log "Mount of $dev on $MEDIUM_DIR failed"
+                return 1
+            fi
+        fi
+        umount "$tmpdir" 2>/dev/null || true
+    fi
+    rmdir "$tmpdir" 2>/dev/null || true
+    return 1
+}
+
+attempt() {
+    log "Scanning for ISO medium..."
+    for dev in $(find_candidates); do
+        log "  Trying $dev ..."
+        if try_mount "$dev"; then
+            local sq="${MEDIUM_DIR}/${SQUASHFS_REL}"
+            log "SUCCESS: squashfs available at $sq ($(du -sh "$sq" | cut -f1))"
+            return 0
+        fi
+    done
+    return 1
+}
+
+if [ "$WAIT_MODE" = "1" ]; then
+    log "Waiting for live medium (press Ctrl+C to abort)..."
+    while true; do
+        if attempt; then
+            exit 0
+        fi
+        log "  Not found — retrying in 5 s (reconnect the disc now)"
+        sleep 5
+    done
+else
+    attempt || die "No ISO medium with ${SQUASHFS_REL} found. Reconnect the disc and re-run, or use --wait."
+fi
Author	SHA1	Message	Date
Michael Chus	5ba72ab315	Add rsync to initramfs for toram progress output live-boot already uses rsync --progress when /bin/rsync exists; without it the copy falls back to silent cp -a. Add rsync to the ISO package list and install an initramfs-tools hook (bee-rsync) that copies the rsync binary + shared libs into the initrd via copy_exec. The hook then rebuilds the initramfs so the change takes effect in the ISO's initrd.img. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-17 23:52:47 +03:00
Michael Chus	63363e9629	Add toram boot entry and Install to RAM resume support - grub.cfg: add "load to RAM (toram)" entry to advanced submenu - install_to_ram.go: resume from existing /dev/shm/bee-live copy if source medium is unavailable after bee-web restart - tasks.go: fix "Recovered after bee-web restart" shown on every run (check j.lines before first append, not after) - bee-install: retry unsquashfs up to 5x with wait-for-remount on source loss; clear error message with bee-remount-medium hint - bee-remount-medium: new script to find and remount live ISO source after USB/CD reconnect; supports --wait polling mode - 9000-bee-setup: chmod +x for bee-install and bee-remount-medium Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-17 23:48:56 +03:00
Mikhail Chusavitin	5285c0d101	Capture per-run IPMI power and GPU telemetry in power benchmark - Sample IPMI loaded_w per single-card calibration and per ramp step instead of averaging over the entire Phase 2; top-level ServerPower uses the final (all-GPU) ramp step value - Add ServerLoadedW/ServerDeltaW to NvidiaPowerBenchGPU and NvidiaPowerBenchStep so external tooling can compare wall power per phase without re-parsing logs - Write gpu-metrics.csv/.html inside each single-XX/ and step-XX/ subdir; aggregate all phases into a top-level gpu-metrics.csv/.html - Write 00-nvidia-smi-q.log at the start of every power run - Add Telemetry (p95 temp/power/fan/clock) to NvidiaPowerBenchGPU in result.json from the converged calibration attempt - Power benchmark page: split "Achieved W" into Single-card W and Multi-GPU W (StablePowerLimitW); derate highlight and status color now reflect the final multi-GPU limit vs nominal - Performance benchmark page: add Status column and per-GPU score color coding (green/yellow/red) based on gpu.Status and OverallStatus Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-17 17:59:58 +03:00
Mikhail Chusavitin	dca4afb8d0	Seed power ramp with single-card TDP limits	2026-04-16 11:43:01 +03:00
Mikhail Chusavitin	b4280941f5	Move NCCL and NVBandwidth into validate mode	2026-04-16 11:02:30 +03:00
Mikhail Chusavitin	f74976ec4c	Use static overlay wallpaper in ISO build	2026-04-16 10:54:03 +03:00
Mikhail Chusavitin	18e24a9aa5	Estimate fan duty from observed RPM maxima	2026-04-16 10:10:18 +03:00
Mikhail Chusavitin	e306250da7	Disable fp64/fp4 in mixed gpu burn	2026-04-16 10:00:03 +03:00
Mikhail Chusavitin	c5b2081ac9	Disable unstable fp4/fp64 benchmark phases	2026-04-16 09:58:02 +03:00
Michael Chus	434528083e	Power bench: compare GPU-reported TDP vs IPMI server power delta - NvidiaPowerBenchResult gains ServerPower *BenchmarkServerPower - RunNvidiaPowerBench samples IPMI idle before Phase 1 and loaded via background goroutine throughout Phase 2 ramp - renderPowerBenchReport: new "Server vs GPU Power Comparison" table with ratio annotation (✓ match / ⚠ minor / ✗ over-report) - renderPowerBenchSummary: server_idle_w, server_loaded_w, server_delta_w, server_reporting_ratio keys Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-16 07:21:02 +03:00