Add per-precision benchmark phases, weighted TOPS scoring, and ECC tracking

- Split steady window into 6 equal slots: fp8/fp16/fp32/fp64/fp4 + combined
- Each precision phase runs bee-gpu-burn with --precision filter so PowerCVPct reflects single-kernel stability (not round-robin artifact)
- Add fp4 support in bee-gpu-stress.c for Blackwell (cc>=100) via existing CUDA_R_4F_E2M1 guard
- Weighted TOPS: fp64×2.0, fp32×1.0, fp16×0.5, fp8×0.25, fp4×0.125
- SyntheticScore = sum of weighted TOPS from per-precision phases
- MixedScore = sum from combined phase; MixedEfficiency = Mixed/Synthetic
- ComputeScore = SyntheticScore × (1 + MixedEfficiency × 0.3)
- ECC volatile counters sampled before/after each phase and overall
- DegradationReasons: ecc_uncorrected_errors, ecc_corrected_errors
- Report: per-precision stability table with ECC columns, methodology section
- Ramp-up history table redesign: GPU indices as columns, runs as rows

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

audit/internal/platform/benchmark.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"math"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"regexp"
 	"sort"
@@ -72,6 +73,11 @@ var (
 	benchmarkIterationsPattern = regexp.MustCompile(`^([a-z0-9_]+)_iterations=(\d+)$`)
 )
 
+// benchmarkPrecisionPhases lists the precision categories run as individual
+// 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{"fp8", "fp16", "fp32", "fp64", "fp4"}
+
 func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts NvidiaBenchmarkOptions, logFunc func(string)) (string, error) {
 	if ctx == nil {
 		ctx = context.Background()
@@ -108,7 +114,11 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		ServerModel:        readServerModel(),
 		BenchmarkProfile:   spec.Name,
 		ParallelGPUs:       opts.ParallelGPUs,
+		RampStep:           opts.RampStep,
+		RampTotal:          opts.RampTotal,
+		RampRunID:          opts.RampRunID,
 		SelectedGPUIndices: append([]int(nil), selected...),
+		HostConfig:         readBenchmarkHostConfig(),
 		Normalization: BenchmarkNormalization{
 			Status: "full",
 		},
@@ -152,8 +162,16 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		}
 	}()
 
+	// Power calibration: run dcgmi targeted_power while sampling nvidia-smi power.
+	// Returns per-GPU p95 power as an honest TDP reference for PowerSustainScore.
+	calibPowerByIndex := runBenchmarkPowerCalibration(ctx, verboseLog, runDir, selected, logFunc)
+
+	// Start background CPU load sampler — samples every 10s during GPU phases.
+	cpuStopCh := make(chan struct{})
+	cpuSamplesCh := startCPULoadSampler(cpuStopCh, 10)
+
 	if opts.ParallelGPUs {
-		runNvidiaBenchmarkParallel(ctx, verboseLog, runDir, selected, infoByIndex, opts, spec, logFunc, &result, &serverIdleW, &serverLoadedWSum, &serverIdleOK, &serverLoadedOK, &serverLoadedSamples)
+		runNvidiaBenchmarkParallel(ctx, verboseLog, runDir, selected, infoByIndex, opts, spec, logFunc, &result, calibPowerByIndex, &serverIdleW, &serverLoadedWSum, &serverIdleOK, &serverLoadedOK, &serverLoadedSamples)
 	} else {
 
 	for _, idx := range selected {
@@ -173,6 +191,9 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 			gpuResult.BaseGraphicsClockMHz = info.BaseGraphicsClockMHz
 			gpuResult.MaxMemoryClockMHz = info.MaxMemoryClockMHz
 		}
+		if w, ok := calibPowerByIndex[idx]; ok && w > 0 {
+			gpuResult.CalibratedPeakPowerW = w
+		}
 		if norm := findBenchmarkNormalization(result.Normalization.GPUs, idx); norm != nil {
 			gpuResult.LockedGraphicsClockMHz = norm.GPUClockLockMHz
 			gpuResult.LockedMemoryClockMHz = norm.MemoryClockLockMHz
@@ -209,14 +230,56 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 			continue
 		}
 
+		// ── Per-precision stability phases ────────────────────────────────────────
+		// Run each precision category alone so PowerCVPct reflects genuine GPU
+		// power stability, not kernel-mix variance.
+		// Time budget: each phase gets steadySec/numPhases, minimum 60 s.
+		// SteadySec is split equally across all precision phases + 1 combined slot.
+		// Skipped phases (unsupported precision) are simply omitted; combined is fixed.
+		totalSlots := len(benchmarkPrecisionPhases) + 1
+		perPhaseSec := spec.SteadySec / totalSlots
+		if perPhaseSec < 60 {
+			perPhaseSec = 60
+		}
+		eccBase, _ := queryECCCounters(idx)
+		for _, prec := range benchmarkPrecisionPhases {
+			phaseCmd := []string{
+				"bee-gpu-burn",
+				"--seconds", strconv.Itoa(perPhaseSec),
+				"--size-mb", strconv.Itoa(opts.SizeMB),
+				"--devices", strconv.Itoa(idx),
+				"--precision", prec,
+			}
+			logFunc(fmt.Sprintf("GPU %d: %s stability phase (%ds)", idx, prec, perPhaseSec))
+			phaseLogName := fmt.Sprintf("gpu-%d-steady-%s", idx, prec)
+			eccBefore, _ := queryECCCounters(idx)
+			phaseOut, phaseRows, phaseErr := runBenchmarkCommandWithMetrics(ctx, verboseLog, phaseLogName+".log", phaseCmd, nil, []int{idx}, runDir, phaseLogName, logFunc)
+			eccAfter, _ := queryECCCounters(idx)
+			if phaseErr != nil || len(phaseRows) == 0 {
+				continue
+			}
+			phase := BenchmarkPrecisionSteadyPhase{
+				Precision: prec,
+				Steady:    summarizeBenchmarkTelemetry(phaseRows),
+				ECC:       diffECCCounters(eccBefore, eccAfter),
+			}
+			for _, p := range parseBenchmarkBurnLog(string(phaseOut)).Profiles {
+				if p.Supported {
+					phase.TeraOpsPerSec += p.TeraOpsPerSec
+					phase.WeightedTeraOpsPerSec += p.WeightedTeraOpsPerSec
+				}
+			}
+			gpuResult.PrecisionSteady = append(gpuResult.PrecisionSteady, phase)
+		}
+
 		beforeThrottle, _ := queryThrottleCounters(idx)
 		steadyCmd := []string{
 			"bee-gpu-burn",
-			"--seconds", strconv.Itoa(spec.SteadySec),
+			"--seconds", strconv.Itoa(perPhaseSec),
 			"--size-mb", strconv.Itoa(opts.SizeMB),
 			"--devices", strconv.Itoa(idx),
 		}
-		logFunc(fmt.Sprintf("GPU %d: steady compute (%ds)", idx, spec.SteadySec))
+		logFunc(fmt.Sprintf("GPU %d: steady compute (combined, %ds)", idx, perPhaseSec))
 
 		// Sample server power via IPMI in parallel with the steady phase.
 		// We collect readings every 5s and average them.
@@ -277,6 +340,9 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 
 		gpuResult.Steady = summarizeBenchmarkTelemetry(steadyRows)
 		gpuResult.Throttle = diffThrottleCounters(beforeThrottle, afterThrottle)
+		if eccFinal, err := queryECCCounters(idx); err == nil {
+			gpuResult.ECC = diffECCCounters(eccBase, eccFinal)
+		}
 
 		cooldownRows, err := collectBenchmarkSamples(ctx, spec.CooldownSec, []int{idx})
 		if err != nil && err != context.Canceled {
@@ -310,6 +376,16 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		}
 	}
 
+	// Stop CPU load sampler and attach results.
+	close(cpuStopCh)
+	if cpuSamples := <-cpuSamplesCh; len(cpuSamples) > 0 {
+		result.CPULoad = summarizeCPULoad(cpuSamples)
+		if result.CPULoad != nil && result.CPULoad.Status != "ok" {
+			logFunc(fmt.Sprintf("host CPU load during benchmark: avg=%.1f%% max=%.1f%% status=%s",
+				result.CPULoad.AvgPct, result.CPULoad.MaxPct, result.CPULoad.Status))
+		}
+	}
+
 	// Compute server power characterization from accumulated IPMI samples.
 	var gpuReportedSumW float64
 	for _, gpu := range result.GPUs {
@@ -321,6 +397,20 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 	}
 	result.ServerPower = characterizeServerPower(serverIdleW, serverLoadedW, gpuReportedSumW, serverIdleOK && serverLoadedOK)
 
+	// Apply server-power penalty when IPMI reports the server delta is much
+	// lower than GPU-reported sum: GPU power telemetry is over-stated, making
+	// CalibratedPeakPowerW and PowerSustainScore unreliable.
+	// Penalty factor scales from 1.0 (ratio ≥ 0.75, no penalty) down to 0.
+	if sp := result.ServerPower; sp != nil && sp.Available && sp.ReportingRatio > 0 && sp.ReportingRatio < 0.75 {
+		factor := sp.ReportingRatio / 0.75
+		for i := range result.GPUs {
+			result.GPUs[i].Scores.CompositeScore *= factor
+			result.GPUs[i].Notes = append(result.GPUs[i].Notes,
+				fmt.Sprintf("server-power penalty applied (reporting_ratio=%.2f < 0.75): composite score reduced to %.1f%%",
+					sp.ReportingRatio, factor*100))
+		}
+	}
+
 	result.Findings = buildBenchmarkFindings(result)
 	result.OverallStatus = benchmarkOverallStatus(result)
 
@@ -421,8 +511,11 @@ func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []b
 
 	// Split the verbose output into per-GPU sections on "^GPU " lines.
 	gpuSectionRe := regexp.MustCompile(`(?m)^GPU\s+([\dA-Fa-f:\.]+)`)
-	maxGfxRe := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Graphics\s*:\s*(\d+)\s*MHz`)
-	maxMemRe := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Memory\s*:\s*(\d+)\s*MHz`)
+	maxGfxRe      := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Graphics\s*:\s*(\d+)\s*MHz`)
+	maxMemRe      := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Memory\s*:\s*(\d+)\s*MHz`)
+	defaultPwrRe  := regexp.MustCompile(`(?i)Default Power Limit\s*:\s*([0-9.]+)\s*W`)
+	currentPwrRe  := regexp.MustCompile(`(?i)Current Power Limit\s*:\s*([0-9.]+)\s*W`)
+	smCountRe     := regexp.MustCompile(`(?i)Multiprocessor Count\s*:\s*(\d+)`)
 
 	sectionStarts := gpuSectionRe.FindAllSubmatchIndex(nvsmiQ, -1)
 	for i, loc := range sectionStarts {
@@ -443,17 +536,14 @@ func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []b
 			continue
 		}
 
-		info := infoByIndex[benchIdx]
-		if info.MaxGraphicsClockMHz > 0 && info.MaxMemoryClockMHz > 0 {
-			continue // already populated
-		}
-
 		end := len(nvsmiQ)
 		if i+1 < len(sectionStarts) {
 			end = sectionStarts[i+1][0]
 		}
 		section := nvsmiQ[loc[0]:end]
 
+		info := infoByIndex[benchIdx]
+
 		if info.MaxGraphicsClockMHz == 0 {
 			if m := maxGfxRe.FindSubmatch(section); m != nil {
 				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil {
@@ -468,6 +558,27 @@ func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []b
 				}
 			}
 		}
+		if info.DefaultPowerLimitW == 0 {
+			if m := defaultPwrRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil && v > 0 {
+					info.DefaultPowerLimitW = v
+				}
+			}
+		}
+		if info.PowerLimitW == 0 {
+			if m := currentPwrRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil && v > 0 {
+					info.PowerLimitW = v
+				}
+			}
+		}
+		if info.MultiprocessorCount == 0 {
+			if m := smCountRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.Atoi(string(m[1])); err == nil && v > 0 {
+					info.MultiprocessorCount = v
+				}
+			}
+		}
 		infoByIndex[benchIdx] = info
 	}
 }
@@ -750,8 +861,11 @@ func parseBenchmarkBurnLog(raw string) benchmarkBurnParseResult {
 			Iterations: profile.iterations,
 			Notes:      profile.notes,
 		}
+		w := precisionWeight(profile.category)
+		precision.Weight = w
 		if profile.supported && result.DurationSec > 0 && profile.m > 0 && profile.n > 0 && profile.k > 0 && profile.iterations > 0 {
 			precision.TeraOpsPerSec = (2.0 * float64(profile.m) * float64(profile.n) * float64(profile.k) * float64(profile.iterations)) / float64(result.DurationSec) / 1e12
+			precision.WeightedTeraOpsPerSec = precision.TeraOpsPerSec * w
 		}
 		result.Profiles = append(result.Profiles, precision)
 	}
@@ -780,6 +894,33 @@ func ensureBenchmarkProfile(profiles map[string]*benchmarkBurnProfile, name stri
 	return profile
 }
 
+// precisionWeight returns the fp32-equivalence factor for a precision category.
+// Each factor represents how much "real" numeric work one operation of that
+// type performs relative to fp32 (single precision = 1.0 baseline):
+//   fp64  = 2.0  — double precision, 2× more bits per operand
+//   fp32  = 1.0  — single precision baseline
+//   fp16  = 0.5  — half precision
+//   fp8   = 0.25 — quarter precision
+//   fp4   = 0.125 — eighth precision
+// Multiplying raw TOPS by the weight gives fp32-equivalent TOPS, enabling
+// cross-precision comparison on the same numeric scale.
+func precisionWeight(category string) float64 {
+	switch category {
+	case "fp64":
+		return 2.0
+	case "fp32_tf32":
+		return 1.0
+	case "fp16_bf16":
+		return 0.5
+	case "fp8":
+		return 0.25
+	case "fp4":
+		return 0.125
+	default:
+		return 1.0
+	}
+}
+
 func stripBenchmarkPrefix(line string) string {
 	if strings.HasPrefix(line, "[gpu ") {
 		if idx := strings.Index(line, "] "); idx >= 0 {
@@ -829,24 +970,72 @@ func summarizeBenchmarkTelemetry(rows []GPUMetricRow) BenchmarkTelemetrySummary
 
 func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	score := BenchmarkScorecard{}
-	for _, precision := range gpu.PrecisionResults {
-		if precision.Supported {
-			score.ComputeScore += precision.TeraOpsPerSec
+
+	// 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 {
+		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 {
+			score.MixedScore += p.WeightedTeraOpsPerSec
 		}
 	}
-	// Use default power limit for sustain score so a manually reduced limit
-	// does not inflate the score. Fall back to enforced limit if default unknown.
-	referencePowerW := gpu.DefaultPowerLimitW
-	if referencePowerW <= 0 {
-		referencePowerW = gpu.PowerLimitW
+
+	// MixedEfficiency = MixedScore / SyntheticScore.
+	// Measures how well the GPU sustains throughput under concurrent mixed load.
+	// A healthy GPU scores ~0.8–0.95; severe degradation suggests bandwidth
+	// contention or scheduler inefficiency.
+	if score.SyntheticScore > 0 && score.MixedScore > 0 {
+		score.MixedEfficiency = score.MixedScore / score.SyntheticScore
 	}
-	if referencePowerW > 0 {
-		score.PowerSustainScore = math.Min(100, (gpu.Steady.AvgPowerW/referencePowerW)*100)
+
+	// ComputeScore = SyntheticScore × (1 + MixedEfficiency × 0.3).
+	// SyntheticScore is the primary signal; MixedEfficiency adds up to +30%
+	// bonus for GPUs that handle mixed-precision concurrency well.
+	// Falls back to MixedScore alone when per-precision data is absent.
+	switch {
+	case score.SyntheticScore > 0:
+		score.ComputeScore = score.SyntheticScore * (1 + score.MixedEfficiency*0.3)
+	case score.MixedScore > 0:
+		score.ComputeScore = score.MixedScore
+	}
+	// PowerSustainScore: measures how close the GPU came to its rated TDP under
+	// a full-spectrum load (dcgmi targeted_power). 100 = exactly at rated TDP.
+	// Penalty applied symmetrically for both under- and over-TDP deviations:
+	//   score = max(0, 100 − |measured − rated| / rated × 100)
+	// Under-TDP → power delivery / cooling issue.
+	// Over-TDP  → power limit not properly enforced / power regulation fault.
+	// Falls back to 0 if calibration was not performed (dcgmi unavailable).
+	{
+		ref := gpu.DefaultPowerLimitW
+		if ref <= 0 {
+			ref = gpu.PowerLimitW
+		}
+		if gpu.CalibratedPeakPowerW > 0 && ref > 0 {
+			deviationPct := math.Abs(gpu.CalibratedPeakPowerW-ref) / ref * 100
+			score.PowerSustainScore = clampScore(100 - deviationPct)
+		}
 	}
 	runtimeUS := math.Max(1, gpu.Steady.DurationSec*1e6)
 	thermalRatio := float64(gpu.Throttle.HWThermalSlowdownUS+gpu.Throttle.SWThermalSlowdownUS) / runtimeUS
 	score.ThermalSustainScore = clampScore(100 - thermalRatio*100)
-	score.StabilityScore = clampScore(100 - (gpu.Steady.ClockCVPct*4 + gpu.Steady.PowerCVPct*2 + gpu.Steady.ClockDriftPct*2))
+	// StabilityScore: prefer per-precision steady phases where each window runs a
+	// single kernel type so PowerCVPct is a genuine stability signal (not a
+	// workload-mix artifact). Fall back to combined steady using clock-only metrics
+	// when per-precision data is absent (older results, short profiles).
+	if len(gpu.PrecisionSteady) > 0 {
+		var sum float64
+		for _, p := range gpu.PrecisionSteady {
+			sum += clampScore(100 - (p.Steady.ClockCVPct*4 + p.Steady.PowerCVPct*2 + p.Steady.ClockDriftPct*2))
+		}
+		score.StabilityScore = sum / float64(len(gpu.PrecisionSteady))
+	} else {
+		score.StabilityScore = clampScore(100 - (gpu.Steady.ClockCVPct*4 + gpu.Steady.ClockDriftPct*2))
+	}
 	score.CompositeScore = compositeBenchmarkScore(score)
 	if gpu.MultiprocessorCount > 0 && gpu.Steady.AvgGraphicsClockMHz > 0 && score.ComputeScore > 0 {
 		score.TOPSPerSMPerGHz = score.ComputeScore / float64(gpu.MultiprocessorCount) / (gpu.Steady.AvgGraphicsClockMHz / 1000.0)
@@ -855,7 +1044,15 @@ func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 }
 
 func compositeBenchmarkScore(score BenchmarkScorecard) float64 {
-	quality := 0.40 + 0.20*(score.PowerSustainScore/100.0) + 0.20*(score.ThermalSustainScore/100.0) + 0.20*(score.StabilityScore/100.0)
+	// Weights after introducing calibrated power reference:
+	//   base        0.35 — floor so a GPU that fails all sustain checks still scores
+	//   thermal     0.25 — heaviest: throttle counters are the most reliable signal
+	//   stability   0.25 — clock/power variance matters for reproducibility
+	//   power       0.15 — GPU reaches rated TDP under targeted_power? lower weight
+	//                       because calibration may be absent (dcgmi not installed)
+	//   NCCL bonus  0.10 — interconnect health
+	//   cap         1.10
+	quality := 0.35 + 0.15*(score.PowerSustainScore/100.0) + 0.25*(score.ThermalSustainScore/100.0) + 0.25*(score.StabilityScore/100.0)
 	if score.InterconnectScore > 0 {
 		quality += 0.10
 	}
@@ -886,6 +1083,12 @@ func detectBenchmarkDegradationReasons(gpu BenchmarkGPUResult, normalizationStat
 	if normalizationStatus != "full" {
 		reasons = append(reasons, "normalization_partial")
 	}
+	if gpu.ECC.Uncorrected > 0 {
+		reasons = append(reasons, "ecc_uncorrected_errors")
+	}
+	if gpu.ECC.Corrected > 0 {
+		reasons = append(reasons, "ecc_corrected_errors")
+	}
 	return dedupeStrings(reasons)
 }
 
@@ -987,6 +1190,36 @@ func diffThrottleCounters(before, after BenchmarkThrottleCounters) BenchmarkThro
 	}
 }
 
+func queryECCCounters(gpuIndex int) (BenchmarkECCCounters, error) {
+	out, err := satExecCommand(
+		"nvidia-smi",
+		"--id="+strconv.Itoa(gpuIndex),
+		"--query-gpu=ecc.errors.corrected.volatile.total,ecc.errors.uncorrected.volatile.total",
+		"--format=csv,noheader,nounits",
+	).Output()
+	if err != nil {
+		return BenchmarkECCCounters{}, err
+	}
+	fields := strings.Split(strings.TrimSpace(string(out)), ",")
+	if len(fields) < 2 {
+		return BenchmarkECCCounters{}, fmt.Errorf("unexpected ECC counter columns: %q", strings.TrimSpace(string(out)))
+	}
+	corrected, err1 := strconv.ParseUint(strings.TrimSpace(fields[0]), 10, 64)
+	uncorrected, err2 := strconv.ParseUint(strings.TrimSpace(fields[1]), 10, 64)
+	if err1 != nil || err2 != nil {
+		// ECC may be disabled on this GPU — return zero counters silently.
+		return BenchmarkECCCounters{}, nil
+	}
+	return BenchmarkECCCounters{Corrected: corrected, Uncorrected: uncorrected}, nil
+}
+
+func diffECCCounters(before, after BenchmarkECCCounters) BenchmarkECCCounters {
+	return BenchmarkECCCounters{
+		Corrected:   saturatingSub(after.Corrected, before.Corrected),
+		Uncorrected: saturatingSub(after.Uncorrected, before.Uncorrected),
+	}
+}
+
 func queryActiveComputeApps(gpuIndices []int) ([]string, error) {
 	args := []string{
 		"--query-compute-apps=gpu_uuid,pid,process_name",
@@ -1064,6 +1297,10 @@ func buildBenchmarkFindings(result NvidiaBenchmarkResult) []string {
 				findings = append(findings, fmt.Sprintf("GPU %d showed unstable clocks/power over the benchmark window.", gpu.Index))
 			case "normalization_partial":
 				findings = append(findings, fmt.Sprintf("GPU %d ran without full benchmark normalization.", gpu.Index))
+			case "ecc_uncorrected_errors":
+				findings = append(findings, fmt.Sprintf("GPU %d reported %d uncorrected ECC error(s) — possible hardware fault.", gpu.Index, gpu.ECC.Uncorrected))
+			case "ecc_corrected_errors":
+				findings = append(findings, fmt.Sprintf("GPU %d reported %d corrected ECC error(s) — possible DRAM degradation.", gpu.Index, gpu.ECC.Corrected))
 			}
 		}
 		if gpu.Backend == "driver-ptx" {
@@ -1075,16 +1312,57 @@ func buildBenchmarkFindings(result NvidiaBenchmarkResult) []string {
 				gpu.Index, gpu.PowerLimitW, gpu.DefaultPowerLimitW, gpu.PowerLimitW/gpu.DefaultPowerLimitW*100,
 			))
 		}
+		// Flag significant TDP deviation (over or under) from calibration.
+		if gpu.CalibratedPeakPowerW > 0 {
+			ref := gpu.DefaultPowerLimitW
+			if ref <= 0 {
+				ref = gpu.PowerLimitW
+			}
+			if ref > 0 {
+				deviationPct := (gpu.CalibratedPeakPowerW - ref) / ref * 100
+				switch {
+				case deviationPct < -10:
+					findings = append(findings, fmt.Sprintf(
+						"GPU %d reached only %.0f W (%.0f%% of rated %.0f W) under targeted_power. Check power delivery or cooling.",
+						gpu.Index, gpu.CalibratedPeakPowerW, gpu.CalibratedPeakPowerW/ref*100, ref,
+					))
+				case deviationPct > 5:
+					findings = append(findings, fmt.Sprintf(
+						"GPU %d exceeded rated TDP: %.0f W measured vs %.0f W rated (+%.0f%%). Power limit may not be enforced correctly.",
+						gpu.Index, gpu.CalibratedPeakPowerW, ref, deviationPct,
+					))
+				}
+			}
+		}
 	}
 	if result.Interconnect != nil && result.Interconnect.Supported {
 		findings = append(findings, fmt.Sprintf("Multi-GPU all_reduce max bus bandwidth: %.1f GB/s.", result.Interconnect.MaxBusBWGBps))
 	}
+	if cl := result.CPULoad; cl != nil {
+		switch cl.Status {
+		case "high":
+			findings = append(findings, fmt.Sprintf(
+				"Host CPU load was elevated during the benchmark (avg %.1f%%, max %.1f%%). A competing CPU workload may skew GPU results.",
+				cl.AvgPct, cl.MaxPct,
+			))
+		case "unstable":
+			findings = append(findings, fmt.Sprintf(
+				"Host CPU load was erratic during the benchmark (avg %.1f%%, p95 %.1f%%). Results may be less reproducible.",
+				cl.AvgPct, cl.P95Pct,
+			))
+		}
+	}
 	if sp := result.ServerPower; sp != nil && sp.Available && sp.GPUReportedSumW > 0 {
 		if sp.ReportingRatio < 0.75 {
 			findings = append(findings, fmt.Sprintf(
-				"GPU power reporting may be unreliable: server delta %.0f W vs GPU-reported %.0f W (ratio %.2f). GPU telemetry likely over-reports actual consumption.",
+				"GPU power reporting may be unreliable: server delta %.0f W vs GPU-reported %.0f W (ratio %.2f). GPU telemetry likely over-reports actual consumption. Composite scores have been penalized accordingly.",
 				sp.DeltaW, sp.GPUReportedSumW, sp.ReportingRatio,
 			))
+		} else if sp.ReportingRatio > 1.25 {
+			findings = append(findings, fmt.Sprintf(
+				"Server power delta %.0f W exceeds GPU-reported sum %.0f W by %.0f%%. Other components (CPU, NVMe, networking) may be drawing substantial power under GPU load.",
+				sp.DeltaW, sp.GPUReportedSumW, (sp.ReportingRatio-1)*100,
+			))
 		}
 	}
 	return dedupeStrings(findings)
@@ -1389,6 +1667,7 @@ func runNvidiaBenchmarkParallel(
 	spec benchmarkProfileSpec,
 	logFunc func(string),
 	result *NvidiaBenchmarkResult,
+	calibPowerByIndex map[int]float64,
 	serverIdleW *float64, serverLoadedWSum *float64,
 	serverIdleOK *bool, serverLoadedOK *bool, serverLoadedSamples *int,
 ) {
@@ -1410,6 +1689,9 @@ func runNvidiaBenchmarkParallel(
 			r.BaseGraphicsClockMHz = info.BaseGraphicsClockMHz
 			r.MaxMemoryClockMHz = info.MaxMemoryClockMHz
 		}
+		if w, ok := calibPowerByIndex[idx]; ok && w > 0 {
+			r.CalibratedPeakPowerW = w
+		}
 		if norm := findBenchmarkNormalization(result.Normalization.GPUs, idx); norm != nil {
 			r.LockedGraphicsClockMHz = norm.GPUClockLockMHz
 			r.LockedMemoryClockMHz = norm.MemoryClockLockMHz
@@ -1458,20 +1740,75 @@ func runNvidiaBenchmarkParallel(
 		}
 	}
 
+	// ── Per-precision stability phases (parallel) ─────────────────────────────
+	totalSlots := len(benchmarkPrecisionPhases) + 1
+	perPhaseSec := spec.SteadySec / totalSlots
+	if perPhaseSec < 60 {
+		perPhaseSec = 60
+	}
+	eccBase := make(map[int]BenchmarkECCCounters, len(selected))
+	for _, idx := range selected {
+		eccBase[idx], _ = queryECCCounters(idx)
+	}
+	for _, prec := range benchmarkPrecisionPhases {
+		phaseCmd := []string{
+			"bee-gpu-burn",
+			"--seconds", strconv.Itoa(perPhaseSec),
+			"--size-mb", strconv.Itoa(opts.SizeMB),
+			"--devices", allDevices,
+			"--precision", prec,
+		}
+		logFunc(fmt.Sprintf("GPUs %s: %s stability phase (%ds)", allDevices, prec, perPhaseSec))
+		phaseLogName := "gpu-all-steady-" + prec
+		eccBeforePhase := make(map[int]BenchmarkECCCounters, len(selected))
+		for _, idx := range selected {
+			eccBeforePhase[idx], _ = queryECCCounters(idx)
+		}
+		phaseOut, phaseRows, phaseErr := runBenchmarkCommandWithMetrics(ctx, verboseLog, phaseLogName+".log", phaseCmd, nil, selected, runDir, phaseLogName, logFunc)
+		eccAfterPhase := make(map[int]BenchmarkECCCounters, len(selected))
+		for _, idx := range selected {
+			eccAfterPhase[idx], _ = queryECCCounters(idx)
+		}
+		if phaseErr != nil || len(phaseRows) == 0 {
+			continue
+		}
+		parseByGPU := parseBenchmarkBurnLogByGPU(string(phaseOut))
+		for _, idx := range selected {
+			perGPU := filterRowsByGPU(phaseRows, idx)
+			if len(perGPU) == 0 {
+				continue
+			}
+			phase := BenchmarkPrecisionSteadyPhase{
+				Precision: prec,
+				Steady:    summarizeBenchmarkTelemetry(perGPU),
+				ECC:       diffECCCounters(eccBeforePhase[idx], eccAfterPhase[idx]),
+			}
+			if pr, ok := parseByGPU[idx]; ok {
+				for _, p := range pr.Profiles {
+					if p.Supported {
+						phase.TeraOpsPerSec += p.TeraOpsPerSec
+						phase.WeightedTeraOpsPerSec += p.WeightedTeraOpsPerSec
+					}
+				}
+			}
+			gpuResults[idx].PrecisionSteady = append(gpuResults[idx].PrecisionSteady, phase)
+		}
+	}
+
 	// Snapshot throttle counters before steady.
 	beforeThrottle := make(map[int]BenchmarkThrottleCounters, len(selected))
 	for _, idx := range selected {
 		beforeThrottle[idx], _ = queryThrottleCounters(idx)
 	}
 
-	// Steady: all GPUs simultaneously.
+	// Steady: all GPUs simultaneously (combined). Fixed at one slot = perPhaseSec.
 	steadyCmd := []string{
 		"bee-gpu-burn",
-		"--seconds", strconv.Itoa(spec.SteadySec),
+		"--seconds", strconv.Itoa(perPhaseSec),
 		"--size-mb", strconv.Itoa(opts.SizeMB),
 		"--devices", allDevices,
 	}
-	logFunc(fmt.Sprintf("GPUs %s: parallel steady compute (%ds)", allDevices, spec.SteadySec))
+	logFunc(fmt.Sprintf("GPUs %s: parallel steady compute (combined, %ds)", allDevices, perPhaseSec))
 
 	// Sample server power via IPMI in parallel with steady phase.
 	ipmiStopCh := make(chan struct{})
@@ -1527,6 +1864,9 @@ func runNvidiaBenchmarkParallel(
 		writeBenchmarkMetricsFiles(runDir, fmt.Sprintf("gpu-%d-steady", idx), perGPU)
 		gpuResults[idx].Steady = summarizeBenchmarkTelemetry(perGPU)
 		gpuResults[idx].Throttle = diffThrottleCounters(beforeThrottle[idx], afterThrottle[idx])
+		if eccFinal, err := queryECCCounters(idx); err == nil {
+			gpuResults[idx].ECC = diffECCCounters(eccBase[idx], eccFinal)
+		}
 
 		if pr, ok := parseResults[idx]; ok {
 			gpuResults[idx].ComputeCapability = pr.ComputeCapability
@@ -1571,3 +1911,225 @@ func runNvidiaBenchmarkParallel(
 		result.GPUs = append(result.GPUs, finalizeBenchmarkGPUResult(*r))
 	}
 }
+
+// readBenchmarkHostConfig reads static CPU and memory configuration from
+// /proc/cpuinfo and /proc/meminfo. Returns nil if neither source is readable.
+func readBenchmarkHostConfig() *BenchmarkHostConfig {
+	cfg := &BenchmarkHostConfig{}
+	populated := false
+
+	// Parse /proc/cpuinfo for CPU model, sockets, cores, threads.
+	if data, err := os.ReadFile("/proc/cpuinfo"); err == nil {
+		socketIDs := map[string]struct{}{}
+		coresPerSocket := map[string]int{}
+		var modelName string
+		threads := 0
+		for _, line := range strings.Split(string(data), "\n") {
+			kv := strings.SplitN(line, ":", 2)
+			if len(kv) != 2 {
+				continue
+			}
+			key := strings.TrimSpace(kv[0])
+			val := strings.TrimSpace(kv[1])
+			switch key {
+			case "processor":
+				threads++
+			case "model name":
+				if modelName == "" {
+					modelName = val
+				}
+			case "physical id":
+				socketIDs[val] = struct{}{}
+			case "cpu cores":
+				// Overwrite per-socket core count (last wins per socket, but all
+				// entries for the same socket report the same value).
+				if physLine := ""; physLine == "" {
+					// We accumulate below by treating cpu cores as a per-thread
+					// field; sum by socket requires a two-pass approach. Use the
+					// simpler approximation: totalCores = threads / (threads per core).
+					_ = val
+				}
+			}
+		}
+		// Second pass: per-socket core count.
+		var curSocket string
+		for _, line := range strings.Split(string(data), "\n") {
+			kv := strings.SplitN(line, ":", 2)
+			if len(kv) != 2 {
+				continue
+			}
+			key := strings.TrimSpace(kv[0])
+			val := strings.TrimSpace(kv[1])
+			switch key {
+			case "physical id":
+				curSocket = val
+			case "cpu cores":
+				if curSocket != "" {
+					if _, seen := coresPerSocket[curSocket]; !seen {
+						v, _ := strconv.Atoi(val)
+						coresPerSocket[curSocket] = v
+					}
+				}
+			}
+		}
+		totalCores := 0
+		for _, c := range coresPerSocket {
+			totalCores += c
+		}
+		cfg.CPUModel = modelName
+		cfg.CPUSockets = len(socketIDs)
+		if cfg.CPUSockets == 0 && threads > 0 {
+			cfg.CPUSockets = 1
+		}
+		cfg.CPUCores = totalCores
+		cfg.CPUThreads = threads
+		if modelName != "" || threads > 0 {
+			populated = true
+		}
+	}
+
+	// Parse /proc/meminfo for total physical RAM.
+	if data, err := os.ReadFile("/proc/meminfo"); err == nil {
+		for _, line := range strings.Split(string(data), "\n") {
+			if strings.HasPrefix(line, "MemTotal:") {
+				fields := strings.Fields(line)
+				if len(fields) >= 2 {
+					kb, _ := strconv.ParseUint(fields[1], 10, 64)
+					cfg.MemTotalGiB = float64(kb) / (1024 * 1024)
+					populated = true
+				}
+				break
+			}
+		}
+	}
+
+	if !populated {
+		return nil
+	}
+	return cfg
+}
+
+// startCPULoadSampler starts a goroutine that samples host CPU load every
+// intervalSec seconds until stopCh is closed, then sends the collected
+// samples on the returned channel.
+func startCPULoadSampler(stopCh <-chan struct{}, intervalSec int) <-chan []float64 {
+	ch := make(chan []float64, 1)
+	go func() {
+		var samples []float64
+		ticker := time.NewTicker(time.Duration(intervalSec) * time.Second)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-stopCh:
+				ch <- samples
+				return
+			case <-ticker.C:
+				if pct := sampleCPULoadPct(); pct > 0 {
+					samples = append(samples, pct)
+				}
+			}
+		}
+	}()
+	return ch
+}
+
+// summarizeCPULoad computes stats over sampled CPU load values and assigns
+// a health status.
+func summarizeCPULoad(samples []float64) *BenchmarkCPULoad {
+	if len(samples) == 0 {
+		return nil
+	}
+	sorted := append([]float64(nil), samples...)
+	sort.Float64s(sorted)
+	var sum float64
+	for _, v := range sorted {
+		sum += v
+	}
+	avg := sum / float64(len(sorted))
+	p95 := sorted[int(float64(len(sorted))*0.95)]
+	max := sorted[len(sorted)-1]
+
+	cl := &BenchmarkCPULoad{
+		AvgPct:  math.Round(avg*10) / 10,
+		MaxPct:  math.Round(max*10) / 10,
+		P95Pct:  math.Round(p95*10) / 10,
+		Samples: len(sorted),
+	}
+
+	// Compute standard deviation to detect instability.
+	var variance float64
+	for _, v := range sorted {
+		d := v - avg
+		variance += d * d
+	}
+	stdDev := math.Sqrt(variance / float64(len(sorted)))
+
+	switch {
+	case avg > 20 || max > 40:
+		cl.Status = "high"
+		cl.Note = fmt.Sprintf("avg %.1f%% max %.1f%% — elevated host CPU load may interfere with GPU benchmark results", avg, max)
+	case stdDev > 12:
+		cl.Status = "unstable"
+		cl.Note = fmt.Sprintf("avg %.1f%% stddev %.1f%% — host CPU load was erratic during the benchmark", avg, stdDev)
+	default:
+		cl.Status = "ok"
+	}
+	return cl
+}
+
+// runBenchmarkPowerCalibration runs a short dcgmi targeted_power test while
+// collecting nvidia-smi power samples in parallel. It returns a map from GPU
+// index to p95 observed power (watts), which is used as the reference for
+// PowerSustainScore instead of the hardware default limit.
+//
+// If dcgmi is unavailable or the run fails the function returns an empty map
+// and the caller falls back to DefaultPowerLimitW. The calibration is skipped
+// gracefully — it must never block or fail the main benchmark.
+func runBenchmarkPowerCalibration(
+	ctx context.Context,
+	verboseLog, runDir string,
+	gpuIndices []int,
+	logFunc func(string),
+) map[int]float64 {
+	const calibDurationSec = 45
+
+	// dcgmi must be present.
+	if _, err := exec.LookPath("dcgmi"); err != nil {
+		logFunc("power calibration: dcgmi not found, skipping (will use default power limit)")
+		return map[int]float64{}
+	}
+
+	logFunc(fmt.Sprintf("power calibration: running dcgmi targeted_power for %ds on GPUs %s", calibDurationSec, joinIndexList(gpuIndices)))
+
+	cmd := nvidiaDCGMNamedDiagCommand("targeted_power", calibDurationSec, gpuIndices)
+	out, rows, err := runBenchmarkCommandWithMetrics(ctx, verboseLog, "power-calibration.log", cmd, nil, gpuIndices, runDir, "power-calibration", logFunc)
+	_ = os.WriteFile(filepath.Join(runDir, "power-calibration.log"), out, 0644)
+	if err != nil {
+		logFunc(fmt.Sprintf("power calibration: dcgmi targeted_power failed (%v), skipping", err))
+		return map[int]float64{}
+	}
+
+	// Group rows by GPU index and compute p95 power for each.
+	result := make(map[int]float64, len(gpuIndices))
+	for _, idx := range gpuIndices {
+		perGPU := filterRowsByGPU(rows, idx)
+		if len(perGPU) == 0 {
+			continue
+		}
+		powers := make([]float64, 0, len(perGPU))
+		for _, r := range perGPU {
+			if r.PowerW > 0 {
+				powers = append(powers, r.PowerW)
+			}
+		}
+		if len(powers) == 0 {
+			continue
+		}
+		p95 := benchmarkPercentile(powers, 95)
+		if p95 > 0 {
+			result[idx] = p95
+			logFunc(fmt.Sprintf("power calibration: GPU %d p95=%.0f W (%d samples)", idx, p95, len(powers)))
+		}
+	}
+	return result
+}

audit/internal/platform/benchmark_report.go

@@ -60,9 +60,17 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 	fmt.Fprintf(&b, "**Profile:** %s  \n", result.BenchmarkProfile)
 	fmt.Fprintf(&b, "**App version:** %s  \n", result.BenchmarkVersion)
 	fmt.Fprintf(&b, "**Generated:** %s  \n", result.GeneratedAt.Format("2006-01-02 15:04:05 UTC"))
-	if result.ParallelGPUs {
+	if result.RampStep > 0 && result.RampTotal > 0 {
+		fmt.Fprintf(&b, "**Ramp-up step:** %d of %d  \n", result.RampStep, result.RampTotal)
+		if result.RampRunID != "" {
+			fmt.Fprintf(&b, "**Ramp-up run ID:** %s  \n", result.RampRunID)
+		}
+	} else if result.ParallelGPUs {
 		fmt.Fprintf(&b, "**Mode:** parallel (all GPUs simultaneously)  \n")
 	}
+	if result.ScalabilityScore > 0 {
+		fmt.Fprintf(&b, "**Scalability score:** %.1f%%  \n", result.ScalabilityScore)
+	}
 	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
 	b.WriteString("\n")
 
@@ -83,10 +91,24 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		b.WriteString("\n")
 	}
 
+	// ── Scoring methodology ───────────────────────────────────────────────────
+	b.WriteString("## Scoring Methodology\n\n")
+	b.WriteString("**Compute score** is derived from two phases:\n\n")
+	b.WriteString("- **Synthetic** — each precision type (fp8, fp16, fp32, fp64, fp4) runs alone for a dedicated window. ")
+	b.WriteString("Measures peak throughput with the full GPU dedicated to one kernel type. ")
+	b.WriteString("Each result is normalised to fp32-equivalent TOPS using precision weights: ")
+	b.WriteString("fp64 ×2.0 · fp32 ×1.0 · fp16 ×0.5 · fp8 ×0.25 · fp4 ×0.125.\n")
+	b.WriteString("- **Mixed** — all precision types run simultaneously (combined phase). ")
+	b.WriteString("Reflects real inference workloads where fp8 matrix ops, fp16 attention and fp32 accumulation compete for bandwidth and SM scheduler slots.\n\n")
+	b.WriteString("**Formula:** `Compute = Synthetic × (1 + MixedEfficiency × 0.3)`\n\n")
+	b.WriteString("where `MixedEfficiency = Mixed / Synthetic`. A GPU that sustains 90 % throughput under mixed load ")
+	b.WriteString("receives a +27 % bonus over its synthetic score; one that drops to 60 % receives +18 %.\n\n")
+	b.WriteString("**Composite score** = `Compute × quality_factor` where quality factors in power sustain, thermal sustain, stability, and interconnect.\n\n")
+
 	// ── Scorecard table ───────────────────────────────────────────────────────
 	b.WriteString("## Scorecard\n\n")
-	b.WriteString("| GPU | Status | Composite | Compute | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
-	b.WriteString("|-----|--------|-----------|---------|-------------|---------------|-----------------|-----------|-------------|\n")
+	b.WriteString("| GPU | Status | Composite | Compute | Synthetic | Mixed | Mixed Eff. | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
+	b.WriteString("|-----|--------|-----------|---------|-----------|-------|------------|-------------|---------------|-----------------|-----------|-------------|\n")
 	for _, gpu := range result.GPUs {
 		name := strings.TrimSpace(gpu.Name)
 		if name == "" {
@@ -100,11 +122,26 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		if gpu.Scores.TOPSPerSMPerGHz > 0 {
 			topsPerSM = fmt.Sprintf("%.3f", gpu.Scores.TOPSPerSMPerGHz)
 		}
-		fmt.Fprintf(&b, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %.1f | %.1f | %.1f | %s |\n",
+		synthetic := "-"
+		if gpu.Scores.SyntheticScore > 0 {
+			synthetic = fmt.Sprintf("%.2f", gpu.Scores.SyntheticScore)
+		}
+		mixed := "-"
+		if gpu.Scores.MixedScore > 0 {
+			mixed = fmt.Sprintf("%.2f", gpu.Scores.MixedScore)
+		}
+		mixedEff := "-"
+		if gpu.Scores.MixedEfficiency > 0 {
+			mixedEff = fmt.Sprintf("%.1f%%", gpu.Scores.MixedEfficiency*100)
+		}
+		fmt.Fprintf(&b, "| GPU %d %s | %s | **%.2f** | %.2f | %s | %s | %s | %s | %.1f | %.1f | %.1f | %s |\n",
 			gpu.Index, name,
 			gpu.Status,
 			gpu.Scores.CompositeScore,
 			gpu.Scores.ComputeScore,
+			synthetic,
+			mixed,
+			mixedEff,
 			topsPerSM,
 			gpu.Scores.PowerSustainScore,
 			gpu.Scores.ThermalSustainScore,
@@ -154,6 +191,35 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		fmt.Fprintf(&b, "| GPU utilisation | %.1f %% | — |\n", gpu.Steady.AvgUsagePct)
 		b.WriteString("\n")
 
+		// Per-precision stability phases.
+		if len(gpu.PrecisionSteady) > 0 {
+			b.WriteString("**Per-precision stability:**\n\n")
+			b.WriteString("| Precision | Clock CV | Power CV | Clock Drift | ECC corr | ECC uncorr |\n|-----------|----------|----------|-------------|----------|------------|\n")
+			for _, p := range gpu.PrecisionSteady {
+				eccCorr := "—"
+				eccUncorr := "—"
+				if !p.ECC.IsZero() {
+					eccCorr = fmt.Sprintf("%d", p.ECC.Corrected)
+					eccUncorr = fmt.Sprintf("%d", p.ECC.Uncorrected)
+				}
+				fmt.Fprintf(&b, "| %s | %.1f%% | %.1f%% | %.1f%% | %s | %s |\n",
+					p.Precision, p.Steady.ClockCVPct, p.Steady.PowerCVPct, p.Steady.ClockDriftPct,
+					eccCorr, eccUncorr)
+			}
+			b.WriteString("\n")
+		} else {
+			// Legacy: show combined-window variance.
+			fmt.Fprintf(&b, "**Clock/power variance (combined window):** clock CV %.1f%% · power CV %.1f%% · clock drift %.1f%%\n\n",
+				gpu.Steady.ClockCVPct, gpu.Steady.PowerCVPct, gpu.Steady.ClockDriftPct)
+		}
+
+
+		// ECC summary
+		if !gpu.ECC.IsZero() {
+			fmt.Fprintf(&b, "**ECC errors (total):** corrected=%d uncorrected=%d\n\n",
+				gpu.ECC.Corrected, gpu.ECC.Uncorrected)
+		}
+
 		// Throttle
 		throttle := formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec)
 		if throttle != "none" {
@@ -163,12 +229,14 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		// Precision results
 		if len(gpu.PrecisionResults) > 0 {
 			b.WriteString("**Precision results:**\n\n")
-			b.WriteString("| Precision | TOPS | Lanes | Iterations |\n|-----------|------|-------|------------|\n")
+			b.WriteString("| Precision | TOPS (raw) | Weight | TOPS (fp32-eq) | Lanes | Iterations |\n|-----------|------------|--------|----------------|-------|------------|\n")
 			for _, p := range gpu.PrecisionResults {
 				if p.Supported {
-					fmt.Fprintf(&b, "| %s | %.2f | %d | %d |\n", p.Name, p.TeraOpsPerSec, p.Lanes, p.Iterations)
+					weightStr := fmt.Sprintf("×%.3g", p.Weight)
+					fmt.Fprintf(&b, "| %s | %.2f | %s | %.2f | %d | %d |\n",
+						p.Name, p.TeraOpsPerSec, weightStr, p.WeightedTeraOpsPerSec, p.Lanes, p.Iterations)
 				} else {
-					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — |\n", p.Name)
+					fmt.Fprintf(&b, "| %s | — (unsupported) | — | — | — | — |\n", p.Name)
 				}
 			}
 			b.WriteString("\n")

audit/internal/platform/benchmark_types.go

@@ -2,6 +2,29 @@ package platform
 
 import "time"
 
+// BenchmarkHostConfig holds static CPU and memory configuration captured at
+// benchmark start. Useful for correlating results across runs on different hardware.
+type BenchmarkHostConfig struct {
+	CPUModel    string  `json:"cpu_model,omitempty"`
+	CPUSockets  int     `json:"cpu_sockets,omitempty"`
+	CPUCores    int     `json:"cpu_cores,omitempty"`
+	CPUThreads  int     `json:"cpu_threads,omitempty"`
+	MemTotalGiB float64 `json:"mem_total_gib,omitempty"`
+}
+
+// BenchmarkCPULoad summarises host CPU utilisation sampled during the GPU
+// steady-state phase. High or unstable CPU load during a GPU benchmark may
+// indicate a competing workload or a CPU-bound driver bottleneck.
+type BenchmarkCPULoad struct {
+	AvgPct  float64 `json:"avg_pct"`
+	MaxPct  float64 `json:"max_pct"`
+	P95Pct  float64 `json:"p95_pct"`
+	Samples int     `json:"samples"`
+	// Status is "ok", "high", or "unstable".
+	Status string `json:"status"`
+	Note   string `json:"note,omitempty"`
+}
+
 const (
 	NvidiaBenchmarkProfileStandard  = "standard"
 	NvidiaBenchmarkProfileStability = "stability"
@@ -14,7 +37,10 @@ type NvidiaBenchmarkOptions struct {
 	GPUIndices        []int
 	ExcludeGPUIndices []int
 	RunNCCL           bool
-	ParallelGPUs      bool // run all selected GPUs simultaneously instead of sequentially
+	ParallelGPUs      bool   // run all selected GPUs simultaneously instead of sequentially
+	RampStep          int    // 1-based step index within a ramp-up run (0 = not a ramp-up)
+	RampTotal         int    // total number of ramp-up steps in this run
+	RampRunID         string // shared identifier across all steps of the same ramp-up run
 }
 
 
@@ -25,11 +51,17 @@ type NvidiaBenchmarkResult struct {
 	ServerModel        string                       `json:"server_model,omitempty"`
 	BenchmarkProfile   string                       `json:"benchmark_profile"`
 	ParallelGPUs       bool                         `json:"parallel_gpus,omitempty"`
+	RampStep           int                          `json:"ramp_step,omitempty"`
+	RampTotal          int                          `json:"ramp_total,omitempty"`
+	RampRunID          string                       `json:"ramp_run_id,omitempty"`
+	ScalabilityScore   float64                      `json:"scalability_score,omitempty"`
 	OverallStatus      string                       `json:"overall_status"`
 	SelectedGPUIndices []int                        `json:"selected_gpu_indices"`
 	Findings           []string                     `json:"findings,omitempty"`
 	Warnings           []string                     `json:"warnings,omitempty"`
 	Normalization      BenchmarkNormalization       `json:"normalization"`
+	HostConfig         *BenchmarkHostConfig         `json:"host_config,omitempty"`
+	CPULoad            *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
 	GPUs               []BenchmarkGPUResult         `json:"gpus"`
 	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
 	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`
@@ -63,16 +95,24 @@ type BenchmarkGPUResult struct {
 	PowerLimitW            float64                    `json:"power_limit_w,omitempty"`
 	MultiprocessorCount    int                        `json:"multiprocessor_count,omitempty"`
 	DefaultPowerLimitW     float64                    `json:"default_power_limit_w,omitempty"`
+	// CalibratedPeakPowerW is the p95 power measured during a short
+	// dcgmi targeted_power calibration run before the main benchmark.
+	// Used as the reference denominator for PowerSustainScore instead of
+	// the hardware default limit, which bee-gpu-burn cannot reach.
+	CalibratedPeakPowerW   float64                    `json:"calibrated_peak_power_w,omitempty"`
 	MaxGraphicsClockMHz    float64                    `json:"max_graphics_clock_mhz,omitempty"`
 	BaseGraphicsClockMHz   float64                    `json:"base_graphics_clock_mhz,omitempty"`
 	MaxMemoryClockMHz      float64                    `json:"max_memory_clock_mhz,omitempty"`
 	LockedGraphicsClockMHz float64                    `json:"locked_graphics_clock_mhz,omitempty"`
 	LockedMemoryClockMHz   float64                    `json:"locked_memory_clock_mhz,omitempty"`
-	Baseline               BenchmarkTelemetrySummary  `json:"baseline"`
-	Steady                 BenchmarkTelemetrySummary  `json:"steady"`
-	Cooldown               BenchmarkTelemetrySummary  `json:"cooldown"`
-	Throttle               BenchmarkThrottleCounters  `json:"throttle_counters"`
-	PrecisionResults       []BenchmarkPrecisionResult `json:"precision_results,omitempty"`
+	Baseline               BenchmarkTelemetrySummary       `json:"baseline"`
+	Steady                 BenchmarkTelemetrySummary       `json:"steady"`
+	PrecisionSteady        []BenchmarkPrecisionSteadyPhase `json:"precision_steady,omitempty"`
+	Cooldown               BenchmarkTelemetrySummary       `json:"cooldown"`
+	Throttle               BenchmarkThrottleCounters       `json:"throttle_counters"`
+	// ECC error delta accumulated over the full benchmark (all phases combined).
+	ECC                    BenchmarkECCCounters            `json:"ecc,omitempty"`
+	PrecisionResults       []BenchmarkPrecisionResult      `json:"precision_results,omitempty"`
 	Scores                 BenchmarkScorecard         `json:"scores"`
 	DegradationReasons     []string                   `json:"degradation_reasons,omitempty"`
 	Notes                  []string                   `json:"notes,omitempty"`
@@ -105,6 +145,18 @@ type BenchmarkThrottleCounters struct {
 	HWPowerBrakeSlowdownUS uint64 `json:"hw_power_brake_slowdown_us"`
 }
 
+// BenchmarkECCCounters holds ECC error counts sampled at a point in time.
+// Corrected = single-bit errors fixed by ECC (DRAM degradation).
+// Uncorrected = double-bit errors that could not be corrected (serious fault).
+// Both are volatile (since last driver reset), not persistent.
+type BenchmarkECCCounters struct {
+	Corrected   uint64 `json:"corrected"`
+	Uncorrected uint64 `json:"uncorrected"`
+}
+
+func (e BenchmarkECCCounters) Total() uint64 { return e.Corrected + e.Uncorrected }
+func (e BenchmarkECCCounters) IsZero() bool  { return e.Corrected == 0 && e.Uncorrected == 0 }
+
 type BenchmarkPrecisionResult struct {
 	Name          string  `json:"name"`
 	Category      string  `json:"category"`
@@ -115,19 +167,31 @@ type BenchmarkPrecisionResult struct {
 	K             uint64  `json:"k,omitempty"`
 	Iterations    uint64  `json:"iterations,omitempty"`
 	TeraOpsPerSec float64 `json:"teraops_per_sec,omitempty"`
+	// Weight is the fp32-equivalence factor for this precision category.
+	// fp32 = 1.0 (baseline), fp64 = 2.0, fp16 = 0.5, fp8 = 0.25, fp4 = 0.125.
+	// WeightedTOPS = TeraOpsPerSec * Weight gives fp32-equivalent throughput.
+	Weight              float64 `json:"weight,omitempty"`
+	WeightedTeraOpsPerSec float64 `json:"weighted_teraops_per_sec,omitempty"`
 	Notes         string  `json:"notes,omitempty"`
 }
 
 type BenchmarkScorecard struct {
 	ComputeScore        float64 `json:"compute_score"`
+	// SyntheticScore is the sum of fp32-equivalent TOPS from per-precision
+	// steady phases (each precision ran alone, full GPU dedicated).
+	SyntheticScore      float64 `json:"synthetic_score,omitempty"`
+	// MixedScore is the sum of fp32-equivalent TOPS from the combined phase
+	// (all precisions competing simultaneously — closer to real workloads).
+	MixedScore          float64 `json:"mixed_score,omitempty"`
+	// MixedEfficiency = MixedScore / SyntheticScore. Measures how well the GPU
+	// sustains throughput under concurrent mixed-precision load.
+	MixedEfficiency     float64 `json:"mixed_efficiency,omitempty"`
 	PowerSustainScore   float64 `json:"power_sustain_score"`
 	ThermalSustainScore float64 `json:"thermal_sustain_score"`
 	StabilityScore      float64 `json:"stability_score"`
 	InterconnectScore   float64 `json:"interconnect_score"`
 	CompositeScore      float64 `json:"composite_score"`
 	// TOPSPerSMPerGHz is compute efficiency independent of clock speed and SM count.
-	// Comparable across throttle levels and GPU generations. Low value at normal
-	// clocks indicates silicon degradation.
 	TOPSPerSMPerGHz float64 `json:"tops_per_sm_per_ghz,omitempty"`
 }
 
@@ -145,6 +209,20 @@ type BenchmarkServerPower struct {
 	Notes           []string `json:"notes,omitempty"`
 }
 
+// BenchmarkPrecisionSteadyPhase holds per-precision-category telemetry collected
+// during a dedicated single-precision steady window.  Because only one kernel
+// type runs at a time the PowerCVPct here is a genuine stability signal.
+type BenchmarkPrecisionSteadyPhase struct {
+	Precision             string                    `json:"precision"` // e.g. "fp8", "fp16", "fp32"
+	Steady                BenchmarkTelemetrySummary `json:"steady"`
+	TeraOpsPerSec         float64                   `json:"teraops_per_sec,omitempty"`
+	WeightedTeraOpsPerSec float64                   `json:"weighted_teraops_per_sec,omitempty"`
+	// ECC errors accumulated during this precision phase only.
+	// Non-zero corrected = stress-induced DRAM errors for this kernel type.
+	// Any uncorrected = serious fault triggered by this precision workload.
+	ECC                   BenchmarkECCCounters      `json:"ecc,omitempty"`
+}
+
 type BenchmarkInterconnectResult struct {
 	Status             string   `json:"status"`
 	Attempted          bool     `json:"attempted"`

audit/internal/platform/install_to_ram.go

@@ -14,9 +14,17 @@ import (
 func (s *System) IsLiveMediaInRAM() bool {
 	fsType := mountFSType("/run/live/medium")
 	if fsType == "" {
+		// No medium mount at all — fall back to toram kernel parameter.
 		return toramActive()
 	}
-	return strings.EqualFold(fsType, "tmpfs")
+	if strings.EqualFold(fsType, "tmpfs") {
+		return true
+	}
+	// When RunInstallToRAM copies squashfs to /dev/shm/bee-live but the bind
+	// mount of /run/live/medium fails (common for CD-ROM boots), the medium
+	// fstype still shows the CD-ROM type. Check whether the RAM copy exists.
+	files, _ := filepath.Glob("/dev/shm/bee-live/*.squashfs")
+	return len(files) > 0
 }
 
 func (s *System) LiveBootSource() LiveBootSource {

audit/internal/platform/runtime.go

@@ -244,11 +244,17 @@ func findUSBExportMount() string {
 		if readOnly {
 			continue
 		}
-		// Check USB transport via lsblk on the device
+		// Check USB transport via lsblk on the device (or its parent disk for partitions).
 		if !strings.HasPrefix(device, "/dev/") {
 			continue
 		}
-		if blockDeviceTransport(device) == "usb" {
+		checkDev := device
+		// lsblk only reports TRAN for the whole disk, not for partitions (e.g. /dev/sdc1).
+		// Strip trailing partition digits to get the parent disk name.
+		if trimmed := strings.TrimRight(device, "0123456789"); trimmed != device && len(trimmed) > len("/dev/") {
+			checkDev = trimmed
+		}
+		if blockDeviceTransport(checkDev) == "usb" {
 			return mountPoint
 		}
 	}

audit/internal/webui/api.go

@@ -497,6 +497,7 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 				GPUIndices         []int    `json:"gpu_indices"`
 				ExcludeGPUIndices  []int    `json:"exclude_gpu_indices"`
 				StaggerGPUStart    bool     `json:"stagger_gpu_start"`
+				ParallelGPUs       bool     `json:"parallel_gpus"`
 				Loader             string   `json:"loader"`
 			Profile            string   `json:"profile"`
 			DisplayName        string   `json:"display_name"`
@@ -519,6 +520,7 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 				GPUIndices:         body.GPUIndices,
 				ExcludeGPUIndices:  body.ExcludeGPUIndices,
 				StaggerGPUStart:    body.StaggerGPUStart,
+				ParallelGPUs:       body.ParallelGPUs,
 				Loader:             body.Loader,
 			BurnProfile:        body.Profile,
 			DisplayName:        body.DisplayName,
@@ -571,10 +573,18 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 	if body.RampUp != nil {
 		rampUp = *body.RampUp
 	}
+	// Build a descriptive base name that includes profile and mode so the task
+	// list is self-explanatory without opening individual task detail pages.
+	profile := strings.TrimSpace(body.Profile)
+	if profile == "" {
+		profile = "standard"
+	}
 	name := taskDisplayName("nvidia-benchmark", "", "")
 	if strings.TrimSpace(body.DisplayName) != "" {
 		name = body.DisplayName
 	}
+	// Append profile tag.
+	name = fmt.Sprintf("%s · %s", name, profile)
 
 	if rampUp && len(body.GPUIndices) > 1 {
 		// Ramp-up mode: resolve GPU list, then create one task per prefix
@@ -594,10 +604,11 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 			rampUp = false
 		} else {
 			now := time.Now()
+			rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
 			var allTasks []*Task
 			for step := 1; step <= len(resolved); step++ {
 				subset := resolved[:step]
-				stepName := fmt.Sprintf("%s [ramp %d/%d: GPU %s]", name, step, len(resolved), formatGPUIndexList(subset))
+				stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
 				t := &Task{
 					ID:        newJobID("benchmark-nvidia"),
 					Name:      stepName,
@@ -611,6 +622,9 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 						BenchmarkProfile: body.Profile,
 						RunNCCL:          runNCCL && step == len(resolved),
 						ParallelGPUs:     true,
+						RampStep:         step,
+						RampTotal:        len(resolved),
+						RampRunID:        rampRunID,
 						DisplayName:      stepName,
 					},
 				}
@@ -624,6 +638,13 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 		}
 	}
 
+	// For non-ramp tasks append mode tag.
+	if parallelGPUs {
+		name = fmt.Sprintf("%s · parallel", name)
+	} else {
+		name = fmt.Sprintf("%s · sequential", name)
+	}
+
 	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
 		GPUIndices:        body.GPUIndices,
 		ExcludeGPUIndices: body.ExcludeGPUIndices,

audit/internal/webui/pages.go

@@ -330,6 +330,33 @@ func renderHardwareSummaryCard(opts HandlerOptions) string {
 
 	var b strings.Builder
 	b.WriteString(`<div class="card"><div class="card-head">Hardware Summary</div><div class="card-body">`)
+
+	// Server identity block above the component table.
+	{
+		var model, serial string
+		parts := []string{}
+		if hw.Board.Manufacturer != nil && strings.TrimSpace(*hw.Board.Manufacturer) != "" {
+			parts = append(parts, strings.TrimSpace(*hw.Board.Manufacturer))
+		}
+		if hw.Board.ProductName != nil && strings.TrimSpace(*hw.Board.ProductName) != "" {
+			parts = append(parts, strings.TrimSpace(*hw.Board.ProductName))
+		}
+		if len(parts) > 0 {
+			model = strings.Join(parts, " ")
+		}
+		serial = strings.TrimSpace(hw.Board.SerialNumber)
+		if model != "" || serial != "" {
+			b.WriteString(`<div style="margin-bottom:14px">`)
+			if model != "" {
+				fmt.Fprintf(&b, `<div style="font-size:16px;font-weight:700;margin-bottom:2px">%s</div>`, html.EscapeString(model))
+			}
+			if serial != "" {
+				fmt.Fprintf(&b, `<div style="font-size:12px;color:var(--muted)">S/N: %s</div>`, html.EscapeString(serial))
+			}
+			b.WriteString(`</div>`)
+		}
+	}
+
 	b.WriteString(`<table style="width:auto">`)
 	writeRow := func(label, value, badgeHTML string) {
 		b.WriteString(fmt.Sprintf(`<tr><td style="padding:6px 14px 6px 0;font-weight:700;white-space:nowrap">%s</td><td style="padding:6px 0;color:var(--muted);font-size:13px">%s</td><td style="padding:6px 0 6px 12px">%s</td></tr>`,
@@ -1279,9 +1306,6 @@ func renderValidate(opts HandlerOptions) string {
 	<div class="card" style="margin-bottom:16px">
 	  <div class="card-head">Validate Profile</div>
 	  <div class="card-body validate-profile-body">
-	    <div class="validate-profile-col">
-	      <div class="form-row" style="margin:0"><label>Cycles</label><input type="number" id="sat-cycles" value="1" min="1" max="100" style="width:100%"></div>
-	    </div>
 	    <div class="validate-profile-col">
 	      <div class="form-row" style="margin:12px 0 0"><label>Mode</label></div>
 	      <label class="cb-row"><input type="radio" name="sat-mode" id="sat-mode-validate" value="validate" checked onchange="satModeChanged()"><span>Validate — quick non-destructive check</span></label>
@@ -1331,12 +1355,6 @@ func renderValidate(opts HandlerOptions) string {
       <p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs...</p>
     </div>
     <p id="sat-gpu-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 0">Select at least one NVIDIA GPU to enable NVIDIA validate tasks.</p>
-    <div style="margin-top:10px;padding-top:10px;border-top:1px solid var(--border)">
-      <label class="sat-gpu-row" title="When checked, multi-GPU tests (PSU Pulse, NCCL, NVBandwidth) run on ALL GPUs in the system regardless of the selection above.">
-        <input type="checkbox" id="sat-multi-gpu-all" checked onchange="satUpdateGPUSelectionNote()">
-        <span><strong>Multi-GPU tests</strong> — use all GPUs <span style="font-size:11px;color:var(--muted)">(PSU Pulse, NCCL, NVBandwidth)</span></span>
-      </label>
-    </div>
   </div>
 </div>
 
@@ -1455,10 +1473,6 @@ function satSelectedGPUIndices() {
     .filter(function(v) { return !Number.isNaN(v); })
     .sort(function(a, b) { return a - b; });
 }
-function satMultiGPUAll() {
-  const cb = document.getElementById('sat-multi-gpu-all');
-  return cb ? cb.checked : true;
-}
 function satUpdateGPUSelectionNote() {
   const note = document.getElementById('sat-gpu-selection-note');
   if (!note) return;
@@ -1467,8 +1481,7 @@ function satUpdateGPUSelectionNote() {
     note.textContent = 'Select at least one NVIDIA GPU to enable NVIDIA validate tasks.';
     return;
   }
-  const multiAll = satMultiGPUAll();
-  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '. Multi-GPU tests: ' + (multiAll ? 'all GPUs in system' : 'selected GPUs only') + '.';
+  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '. Multi-GPU tests will use all selected GPUs.';
 }
 function satRenderGPUList(gpus) {
   const root = document.getElementById('sat-gpu-list');
@@ -1582,15 +1595,8 @@ const nvidiaPerGPUTargets = ['nvidia', 'nvidia-targeted-stress', 'nvidia-targete
 // pulse_test and fabric tests run on all selected GPUs simultaneously
 const nvidiaAllGPUTargets = ['nvidia-pulse', 'nvidia-interconnect', 'nvidia-bandwidth'];
 function satAllGPUIndicesForMulti() {
-  // If "Multi-GPU tests — all GPUs" is checked, return all detected GPUs.
-  // Otherwise fall back to the per-GPU selection.
-  if (satMultiGPUAll()) {
-    return loadSatNvidiaGPUs().then(function(gpus) {
-      return gpus.map(function(g) { return Number(g.index); });
-    });
-  }
-  const sel = satSelectedGPUIndices();
-  return Promise.resolve(sel);
+  // Multi-GPU tests always use the current GPU selection.
+  return Promise.resolve(satSelectedGPUIndices());
 }
 function expandSATTarget(target) {
   if (nvidiaAllGPUTargets.indexOf(target) >= 0) {
@@ -1680,7 +1686,7 @@ function runAMDValidateSet() {
   return runNext(0);
 }
 function runAllSAT() {
-  const cycles = Math.max(1, parseInt(document.getElementById('sat-cycles').value)||1);
+  const cycles = 1;
   const status = document.getElementById('sat-all-status');
   status.textContent = 'Enqueuing...';
   const stressOnlyTargets = ['nvidia-targeted-stress', 'nvidia-targeted-power', 'nvidia-pulse', 'nvidia-interconnect', 'nvidia-bandwidth'];
@@ -1922,23 +1928,10 @@ func renderSATCard(id, label, runAction, headerActions, body string) string {
 
 // ── Benchmark ─────────────────────────────────────────────────────────────────
 
-type benchmarkHistoryColumn struct {
-	key      string
-	label    string
-	name     string
-	index    int
-	parallel bool
-}
-
-type benchmarkHistoryCell struct {
-	score   float64
-	present bool
-}
-
 type benchmarkHistoryRun struct {
 	generatedAt time.Time
 	displayTime string
-	cells       map[string]benchmarkHistoryCell
+	gpuScores   map[int]float64 // GPU index → composite score
 }
 
 func renderBenchmark(opts HandlerOptions) string {
@@ -1967,16 +1960,16 @@ func renderBenchmark(opts HandlerOptions) string {
         </div>
       </div>
       <label class="benchmark-cb-row">
-        <input type="checkbox" id="benchmark-ramp-up" checked onchange="benchmarkUpdateSelectionNote()">
-        <span>Ramp-up mode: run 1 GPU, then 2, then 3… up to all selected GPUs (each step is a separate task)</span>
+        <input type="radio" name="benchmark-mode" value="sequential" onchange="benchmarkUpdateSelectionNote()">
+        <span>Sequential — one GPU at a time</span>
       </label>
-      <label class="benchmark-cb-row">
-        <input type="checkbox" id="benchmark-parallel-gpus">
-        <span>Run all selected GPUs simultaneously (parallel mode, ignored in ramp-up)</span>
+      <label class="benchmark-cb-row" id="benchmark-parallel-label">
+        <input type="radio" name="benchmark-mode" value="parallel" onchange="benchmarkUpdateSelectionNote()">
+        <span>Parallel — all selected GPUs simultaneously</span>
       </label>
-      <label class="benchmark-cb-row">
-        <input type="checkbox" id="benchmark-run-nccl" checked>
-        <span>Run multi-GPU interconnect step (NCCL) only on the selected GPUs</span>
+      <label class="benchmark-cb-row" id="benchmark-ramp-label">
+        <input type="radio" name="benchmark-mode" value="ramp-up" checked onchange="benchmarkUpdateSelectionNote()">
+        <span>Ramp-up — 1 GPU → 2 → … → all selected (separate tasks)</span>
       </label>
       <p id="benchmark-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 14px">Select one GPU for single-card benchmarking or several GPUs for a constrained multi-GPU run.</p>
       <button id="benchmark-run-btn" class="btn btn-primary" onclick="runNvidiaBenchmark()" disabled>&#9654; Run Benchmark</button>
@@ -2029,27 +2022,28 @@ function benchmarkSelectedGPUIndices() {
     .sort(function(a, b) { return a - b; });
 }
 
+function benchmarkMode() {
+  const el = document.querySelector('input[name="benchmark-mode"]:checked');
+  return el ? el.value : 'sequential';
+}
+
 function benchmarkUpdateSelectionNote() {
   const selected = benchmarkSelectedGPUIndices();
   const btn = document.getElementById('benchmark-run-btn');
   const note = document.getElementById('benchmark-selection-note');
-  const nccl = document.getElementById('benchmark-run-nccl');
   if (!selected.length) {
     btn.disabled = true;
     note.textContent = 'Select at least one NVIDIA GPU to run the benchmark.';
     return;
   }
   btn.disabled = false;
-  const rampUp = selected.length > 1 && !!document.getElementById('benchmark-ramp-up').checked;
-  if (rampUp) {
-    note.textContent = 'Ramp-up: will spawn ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). NCCL runs on the final step only.';
+  const mode = benchmarkMode();
+  if (mode === 'ramp-up') {
+    note.textContent = 'Ramp-up: ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). NCCL on final step.';
+  } else if (mode === 'parallel') {
+    note.textContent = 'Parallel: all ' + selected.length + ' GPU(s) simultaneously.' + (selected.length > 1 ? ' NCCL included.' : '');
   } else {
-    note.textContent = 'Selected GPUs: ' + selected.join(', ') + '.';
-    if (nccl && nccl.checked && selected.length < 2) {
-      note.textContent += ' NCCL will be skipped because fewer than 2 GPUs are selected.';
-    } else if (nccl && nccl.checked) {
-      note.textContent += ' NCCL interconnect will use only these GPUs.';
-    }
+    note.textContent = 'Sequential: each GPU benchmarked separately.' + (selected.length > 1 ? ' NCCL included on each.' : '');
   }
 }
 
@@ -2067,6 +2061,33 @@ function benchmarkRenderGPUList(gpus) {
       + '<span><strong>GPU ' + gpu.index + '</strong> — ' + gpu.name + mem + '</span>'
       + '</label>';
   }).join('');
+  benchmarkApplyMultiGPUState(gpus.length);
+  benchmarkUpdateSelectionNote();
+}
+
+// Disable radio options that require multiple GPUs when only one is present.
+function benchmarkApplyMultiGPUState(gpuCount) {
+  var multiValues = ['parallel', 'ramp-up'];
+  var radios = document.querySelectorAll('input[name="benchmark-mode"]');
+  radios.forEach(function(el) {
+    var isMulti = multiValues.indexOf(el.value) >= 0;
+    if (gpuCount < 2 && isMulti) {
+      el.disabled = true;
+      if (el.checked) {
+        // fall back to sequential
+        var seq = document.querySelector('input[name="benchmark-mode"][value="sequential"]');
+        if (seq) seq.checked = true;
+      }
+      var label = el.closest('label');
+      if (label) label.style.opacity = '0.4';
+    } else {
+      el.disabled = false;
+      // restore default: ramp-up checked when ≥2 GPUs
+      if (gpuCount >= 2 && el.value === 'ramp-up') el.checked = true;
+      var label = el.closest('label');
+      if (label) label.style.opacity = '';
+    }
+  });
   benchmarkUpdateSelectionNote();
 }
 
@@ -2104,12 +2125,13 @@ function runNvidiaBenchmark() {
     return;
   }
   if (benchmarkES) { benchmarkES.close(); benchmarkES = null; }
-  const rampUp = selected.length > 1 && !!document.getElementById('benchmark-ramp-up').checked;
-  const parallelGPUs = !rampUp && !!document.getElementById('benchmark-parallel-gpus').checked;
+  const mode = benchmarkMode();
+  const rampUp = mode === 'ramp-up' && selected.length > 1;
+  const parallelGPUs = mode === 'parallel';
   const body = {
     profile: document.getElementById('benchmark-profile').value || 'standard',
     gpu_indices: selected,
-    run_nccl: !!document.getElementById('benchmark-run-nccl').checked,
+    run_nccl: selected.length > 1,
     parallel_gpus: parallelGPUs,
     ramp_up: rampUp,
     display_name: 'NVIDIA Benchmark'
@@ -2166,23 +2188,22 @@ function runNvidiaBenchmark() {
   });
 }
 
-document.getElementById('benchmark-run-nccl').addEventListener('change', benchmarkUpdateSelectionNote);
 benchmarkLoadGPUs();
 </script>`
 }
 
 func renderBenchmarkResultsCard(exportDir string) string {
-	columns, runs := loadBenchmarkHistory(exportDir)
+	maxIdx, runs := loadBenchmarkHistory(exportDir)
 	return renderBenchmarkResultsCardFromRuns(
 		"Benchmark Results",
 		"Composite score by saved benchmark run and GPU.",
 		"No saved benchmark runs yet.",
-		columns,
+		maxIdx,
 		runs,
 	)
 }
 
-func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, columns []benchmarkHistoryColumn, runs []benchmarkHistoryRun) string {
+func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, maxGPUIndex int, runs []benchmarkHistoryRun) string {
 	if len(runs) == 0 {
 		return `<div class="card"><div class="card-head">` + html.EscapeString(title) + `</div><div class="card-body"><p style="color:var(--muted);font-size:13px">` + html.EscapeString(emptyMessage) + `</p></div></div>`
 	}
@@ -2192,22 +2213,22 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 		b.WriteString(`<p style="color:var(--muted);font-size:13px;margin-bottom:12px">` + html.EscapeString(description) + `</p>`)
 	}
 	b.WriteString(`<div style="overflow-x:auto">`)
-	b.WriteString(`<table><thead><tr><th>Test</th><th>Time</th>`)
-	for _, col := range columns {
-		b.WriteString(`<th>` + html.EscapeString(col.label) + `</th>`)
+	b.WriteString(`<table><thead><tr><th>Run</th><th>Time</th>`)
+	for i := 0; i <= maxGPUIndex; i++ {
+		b.WriteString(`<th>GPU ` + strconv.Itoa(i) + `</th>`)
 	}
 	b.WriteString(`</tr></thead><tbody>`)
 	for i, run := range runs {
 		b.WriteString(`<tr>`)
 		b.WriteString(`<td>#` + strconv.Itoa(i+1) + `</td>`)
 		b.WriteString(`<td>` + html.EscapeString(run.displayTime) + `</td>`)
-		for _, col := range columns {
-			cell, ok := run.cells[col.key]
-			if !ok || !cell.present {
+		for idx := 0; idx <= maxGPUIndex; idx++ {
+			score, ok := run.gpuScores[idx]
+			if !ok {
 				b.WriteString(`<td style="color:var(--muted)">-</td>`)
 				continue
 			}
-			b.WriteString(`<td>` + fmt.Sprintf("%.2f", cell.score) + `</td>`)
+			b.WriteString(`<td>` + fmt.Sprintf("%.2f", score) + `</td>`)
 		}
 		b.WriteString(`</tr>`)
 	}
@@ -2215,22 +2236,22 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 	return b.String()
 }
 
-func loadBenchmarkHistory(exportDir string) ([]benchmarkHistoryColumn, []benchmarkHistoryRun) {
+func loadBenchmarkHistory(exportDir string) (int, []benchmarkHistoryRun) {
 	baseDir := app.DefaultBenchmarkBaseDir
 	if strings.TrimSpace(exportDir) != "" {
 		baseDir = filepath.Join(exportDir, "bee-benchmark")
 	}
 	paths, err := filepath.Glob(filepath.Join(baseDir, "gpu-benchmark-*", "result.json"))
 	if err != nil || len(paths) == 0 {
-		return nil, nil
+		return -1, nil
 	}
 	sort.Strings(paths)
 	return loadBenchmarkHistoryFromPaths(paths)
 }
 
-func loadBenchmarkHistoryFromPaths(paths []string) ([]benchmarkHistoryColumn, []benchmarkHistoryRun) {
-	columnByKey := make(map[string]benchmarkHistoryColumn)
+func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun) {
 	runs := make([]benchmarkHistoryRun, 0, len(paths))
+	maxGPUIndex := -1
 	for _, path := range paths {
 		raw, err := os.ReadFile(path)
 		if err != nil {
@@ -2243,102 +2264,22 @@ func loadBenchmarkHistoryFromPaths(paths []string) ([]benchmarkHistoryColumn, []
 		run := benchmarkHistoryRun{
 			generatedAt: result.GeneratedAt,
 			displayTime: result.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
-			cells:       make(map[string]benchmarkHistoryCell),
+			gpuScores:   make(map[int]float64),
 		}
-
-		if result.ParallelGPUs {
-			// All GPUs ran simultaneously — one column per server, score = avg composite.
-			gpuModelCount := make(map[string]int)
-			for _, gpu := range result.GPUs {
-				gpuModelCount[strings.TrimSpace(gpu.Name)]++
-			}
-			scoreSum := make(map[string]float64)
-			scoreCnt := make(map[string]int)
-			for _, gpu := range result.GPUs {
-				key := "parallel|" + strings.TrimSpace(result.ServerModel) + "|" + strings.TrimSpace(gpu.Name)
-				scoreSum[key] += gpu.Scores.CompositeScore
-				scoreCnt[key]++
-				count := gpuModelCount[strings.TrimSpace(gpu.Name)]
-				columnByKey[key] = benchmarkHistoryColumn{
-					key:      key,
-					label:    benchmarkHistoryParallelLabel(result.ServerModel, gpu.Name, count),
-					name:     strings.TrimSpace(gpu.Name),
-					index:    -1,
-					parallel: true,
-				}
-			}
-			for key, sum := range scoreSum {
-				run.cells[key] = benchmarkHistoryCell{score: sum / float64(scoreCnt[key]), present: true}
-			}
-		} else {
-			// Each GPU ran independently — one column per GPU index.
-			for _, gpu := range result.GPUs {
-				key := "gpu|" + strings.TrimSpace(result.ServerModel) + "|" + strings.TrimSpace(gpu.Name) + "|" + strconv.Itoa(gpu.Index)
-				columnByKey[key] = benchmarkHistoryColumn{
-					key:      key,
-					label:    benchmarkHistoryPerGPULabel(gpu.Name, gpu.Index),
-					name:     strings.TrimSpace(gpu.Name),
-					index:    gpu.Index,
-					parallel: false,
-				}
-				run.cells[key] = benchmarkHistoryCell{score: gpu.Scores.CompositeScore, present: true}
+		for _, gpu := range result.GPUs {
+			run.gpuScores[gpu.Index] = gpu.Scores.CompositeScore
+			if gpu.Index > maxGPUIndex {
+				maxGPUIndex = gpu.Index
 			}
 		}
 		runs = append(runs, run)
 	}
-
-	columns := make([]benchmarkHistoryColumn, 0, len(columnByKey))
-	for _, col := range columnByKey {
-		columns = append(columns, col)
-	}
-	// Sequential GPU columns first (sorted by GPU index), then parallel server columns.
-	sort.Slice(columns, func(i, j int) bool {
-		if columns[i].parallel != columns[j].parallel {
-			return !columns[i].parallel // sequential first
-		}
-		if columns[i].parallel {
-			li := strings.ToLower(columns[i].label)
-			lj := strings.ToLower(columns[j].label)
-			if li != lj {
-				return li < lj
-			}
-			return columns[i].key < columns[j].key
-		}
-		// Sequential: sort by GPU index, then name.
-		if columns[i].index != columns[j].index {
-			return columns[i].index < columns[j].index
-		}
-		return strings.ToLower(columns[i].name) < strings.ToLower(columns[j].name)
-	})
 	sort.Slice(runs, func(i, j int) bool {
 		return runs[i].generatedAt.After(runs[j].generatedAt)
 	})
-	return columns, runs
+	return maxGPUIndex, runs
 }
 
-// benchmarkHistoryPerGPULabel formats a label for a single-GPU column: "GPU #N — ModelName".
-func benchmarkHistoryPerGPULabel(gpuName string, index int) string {
-	gpuName = strings.TrimSpace(gpuName)
-	if gpuName == "" {
-		gpuName = "Unknown GPU"
-	}
-	return fmt.Sprintf("GPU #%d — %s", index, gpuName)
-}
-
-// benchmarkHistoryParallelLabel formats a label for an all-GPU parallel column:
-// "ServerModel — N× ModelName (All GPUs)" or "N× ModelName (All GPUs)" if no server.
-func benchmarkHistoryParallelLabel(serverModel, gpuName string, count int) string {
-	serverModel = strings.TrimSpace(serverModel)
-	gpuName = strings.TrimSpace(gpuName)
-	if gpuName == "" {
-		gpuName = "Unknown GPU"
-	}
-	gpuPart := fmt.Sprintf("%d× %s (All GPUs)", count, gpuName)
-	if serverModel == "" {
-		return gpuPart
-	}
-	return fmt.Sprintf("%s — %s", serverModel, gpuPart)
-}
 
 // ── Burn ──────────────────────────────────────────────────────────────────────
 
@@ -2382,10 +2323,20 @@ func renderBurn() string {
 	      <p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs...</p>
 	    </div>
 	    <p id="burn-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 0">Select at least one NVIDIA GPU to enable NVIDIA burn recipes.</p>
-	    <label class="cb-row" style="margin-top:10px">
-	      <input type="checkbox" id="burn-stagger-nvidia">
-	      <span>Ramp selected NVIDIA GPUs one by one before the full-load hold. Smoke: +2 min per GPU, then 5 min with all selected GPUs under load. Acceptance: +10 min per GPU, then at least 1 hour with all selected GPUs under load. Overnight: +1 hour per GPU, then at least 1 hour with all selected GPUs under load, capped at 10 hours total.</span>
-	    </label>
+	    <div style="display:flex;flex-direction:column;gap:4px;margin-top:10px">
+	      <label class="cb-row">
+	        <input type="radio" name="burn-nvidia-mode" value="sequential" checked>
+	        <span>Sequential — selected GPUs one at a time</span>
+	      </label>
+	      <label class="cb-row" id="burn-parallel-label">
+	        <input type="radio" name="burn-nvidia-mode" value="parallel">
+	        <span>Parallel — all selected GPUs simultaneously</span>
+	      </label>
+	      <label class="cb-row" id="burn-ramp-label">
+	        <input type="radio" name="burn-nvidia-mode" value="ramp-up">
+	        <span>Ramp-up — add one GPU at a time</span>
+	      </label>
+	    </div>
 	  </div>
 	</div>
 
@@ -2461,9 +2412,30 @@ function burnSelectedGPUIndices() {
     .sort(function(a, b) { return a - b; });
 }
 
-function burnUseNvidiaRampUp() {
-  const el = document.getElementById('burn-stagger-nvidia');
-  return !!(el && el.checked);
+function burnNvidiaMode() {
+  const el = document.querySelector('input[name="burn-nvidia-mode"]:checked');
+  return el ? el.value : 'sequential';
+}
+
+function burnApplyMultiGPUState(gpuCount) {
+  var multiValues = ['parallel', 'ramp-up'];
+  var radios = document.querySelectorAll('input[name="burn-nvidia-mode"]');
+  radios.forEach(function(el) {
+    var isMulti = multiValues.indexOf(el.value) >= 0;
+    if (gpuCount < 2 && isMulti) {
+      el.disabled = true;
+      if (el.checked) {
+        var seq = document.querySelector('input[name="burn-nvidia-mode"][value="sequential"]');
+        if (seq) seq.checked = true;
+      }
+      var label = el.closest('label');
+      if (label) label.style.opacity = '0.4';
+    } else {
+      el.disabled = false;
+      var label = el.closest('label');
+      if (label) label.style.opacity = '';
+    }
+  });
 }
 
 function burnUpdateSelectionNote() {
@@ -2490,6 +2462,7 @@ function burnRenderGPUList(gpus) {
       + '<span><strong>GPU ' + gpu.index + '</strong> — ' + gpu.name + mem + '</span>'
       + '</label>';
   }).join('');
+  burnApplyMultiGPUState(gpus.length);
   burnUpdateSelectionNote();
 }
 
@@ -2525,8 +2498,11 @@ function enqueueBurnTask(target, label, extra, useSelectedNvidia) {
       return Promise.reject(new Error('Select at least one NVIDIA GPU.'));
     }
     body.gpu_indices = selected;
-    if (burnUseNvidiaRampUp() && selected.length > 1) {
+    const bMode = burnNvidiaMode();
+    if (bMode === 'ramp-up' && selected.length > 1) {
       body.stagger_gpu_start = true;
+    } else if (bMode === 'parallel' && selected.length > 1) {
+      body.parallel_gpus = true;
     }
   }
   return fetch('/api/sat/' + target + '/run', {

audit/internal/webui/server_test.go

@@ -693,8 +693,8 @@ func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	for _, needle := range []string{
 		`Benchmark Results`,
 		`Composite score by saved benchmark run and GPU.`,
-		`GPU #0 — NVIDIA H100 PCIe`,
-		`GPU #1 — NVIDIA H100 PCIe`,
+		`GPU 0`,
+		`GPU 1`,
 		`#1`,
 		wantTime,
 		`1176.25`,

audit/internal/webui/tasks.go

@@ -126,6 +126,9 @@ type taskParams struct {
 	BenchmarkProfile   string   `json:"benchmark_profile,omitempty"`
 	RunNCCL            bool     `json:"run_nccl,omitempty"`
 	ParallelGPUs       bool     `json:"parallel_gpus,omitempty"`
+	RampStep           int      `json:"ramp_step,omitempty"`
+	RampTotal          int      `json:"ramp_total,omitempty"`
+	RampRunID          string   `json:"ramp_run_id,omitempty"`
 	DisplayName        string   `json:"display_name,omitempty"`
 	Device             string   `json:"device,omitempty"` // for install
 	PlatformComponents []string `json:"platform_components,omitempty"`
@@ -637,6 +640,9 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
 			RunNCCL:           t.params.RunNCCL,
 			ParallelGPUs:      t.params.ParallelGPUs,
+			RampStep:          t.params.RampStep,
+			RampTotal:         t.params.RampTotal,
+			RampRunID:         t.params.RampRunID,
 		}, j.append)
 	case "nvidia-compute":
 		if a == nil {

audit/internal/webui/tasks_test.go

@@ -422,7 +422,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	for _, needle := range []string{
 		`Benchmark Results`,
 		`Composite score for this benchmark task.`,
-		`GPU #0 — NVIDIA H100 PCIe`,
+		`GPU 0`,
 		`1176.25`,
 	} {
 		if !strings.Contains(html, needle) {

iso/builder/bee-gpu-stress.c

@@ -1121,6 +1121,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                                int cc_minor,
                                int seconds,
                                int size_mb,
+                               const char *precision_filter,
                                struct stress_report *report) {
     struct cublaslt_api cublas;
     struct prepared_profile prepared[MAX_STRESS_STREAMS * MAX_CUBLAS_PROFILES];
@@ -1159,7 +1160,8 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     }
 
     for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
-        if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc) {
+        if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc &&
+            (precision_filter == NULL || strcmp(k_profiles[i].block_label, precision_filter) == 0)) {
             planned++;
         }
     }
@@ -1218,6 +1220,13 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                           desc->min_cc);
             continue;
         }
+        if (precision_filter != NULL && strcmp(desc->block_label, precision_filter) != 0) {
+            append_detail(report->details,
+                          sizeof(report->details),
+                          "%s=SKIPPED precision_filter\n",
+                          desc->name);
+            continue;
+        }
         for (int lane = 0; lane < stream_count; lane++) {
             CUstream stream = streams[lane];
             if (prepared_count >= (int)(sizeof(prepared) / sizeof(prepared[0]))) {
@@ -1339,6 +1348,7 @@ int main(int argc, char **argv) {
     int seconds = 5;
     int size_mb = 64;
     int device_index = 0;
+    const char *precision_filter = NULL; /* NULL = all; else block_label to match */
     for (int i = 1; i < argc; i++) {
         if ((strcmp(argv[i], "--seconds") == 0 || strcmp(argv[i], "-t") == 0) && i + 1 < argc) {
             seconds = atoi(argv[++i]);
@@ -1346,8 +1356,12 @@ int main(int argc, char **argv) {
             size_mb = atoi(argv[++i]);
         } else if ((strcmp(argv[i], "--device") == 0 || strcmp(argv[i], "-d") == 0) && i + 1 < argc) {
             device_index = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--precision") == 0 && i + 1 < argc) {
+            precision_filter = argv[++i];
         } else {
-            fprintf(stderr, "usage: %s [--seconds N] [--size-mb N] [--device N]\n", argv[0]);
+            fprintf(stderr,
+                    "usage: %s [--seconds N] [--size-mb N] [--device N] [--precision fp8|fp16|fp32|fp64|fp4]\n",
+                    argv[0]);
             return 2;
         }
     }
@@ -1407,7 +1421,7 @@ int main(int argc, char **argv) {
     int ok = 0;
 
 #if HAVE_CUBLASLT_HEADERS
-    ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, &report);
+    ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, precision_filter, &report);
 #endif
     if (!ok) {
         if (!run_ptx_fallback(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, &report)) {

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

@@ -6,10 +6,11 @@ STAGGER_SECONDS=0
 SIZE_MB=0
 DEVICES=""
 EXCLUDE=""
+PRECISION=""
 WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
 
 usage() {
-    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
+    echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3] [--precision fp8|fp16|fp32|fp64|fp4]" >&2
     exit 2
 }
 
@@ -30,6 +31,7 @@ while [ "$#" -gt 0 ]; do
         --size-mb|-m) [ "$#" -ge 2 ] || usage; SIZE_MB="$2"; shift 2 ;;
         --devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
         --exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
+        --precision) [ "$#" -ge 2 ] || usage; PRECISION="$2"; shift 2 ;;
         *) usage ;;
     esac
 done
@@ -88,8 +90,10 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
     extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
     gpu_seconds=$(( SECONDS + extra_sec ))
     echo "starting gpu ${id} size=${gpu_size_mb}MB seconds=${gpu_seconds}"
+    precision_arg=""
+    [ -n "${PRECISION}" ] && precision_arg="--precision ${PRECISION}"
     CUDA_VISIBLE_DEVICES="${id}" \
-        "${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+        "${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" ${precision_arg} >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
     if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then