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

@@ -73,6 +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()
@@ -225,14 +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.
@@ -293,6 +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 {
@@ -811,8 +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)
 	}
@@ -841,6 +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 {
@@ -890,11 +970,39 @@ func summarizeBenchmarkTelemetry(rows []GPUMetricRow) BenchmarkTelemetrySummary
 
 func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	score := BenchmarkScorecard{}
-	for _, precision := range gpu.PrecisionResults {
+
-		if precision.Supported {
+	// SyntheticScore: sum of fp32-equivalent TOPS from per-precision phases.
-			score.ComputeScore += precision.TeraOpsPerSec
+	// 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
 		}
 	}
+
+	// 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
+	}
+
+	// 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:
@@ -915,7 +1023,19 @@ func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	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)
@@ -963,6 +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)
 }
 
@@ -1064,6 +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",
@@ -1141,6 +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" {
@@ -1580,20 +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{})
@@ -1649,6 +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

audit/internal/platform/benchmark_report.go

@@ -91,10 +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("| GPU | Status | Composite | Compute | Synthetic | Mixed | Mixed Eff. | TOPS/SM/GHz | Power Sustain | Thermal Sustain | Stability | Interconnect |\n")
-	b.WriteString("|-----|--------|-----------|---------|-------------|---------------|-----------------|-----------|-------------|\n")
+	b.WriteString("|-----|--------|-----------|---------|-----------|-------|------------|-------------|---------------|-----------------|-----------|-------------|\n")
 	for _, gpu := range result.GPUs {
 		name := strings.TrimSpace(gpu.Name)
 		if name == "" {
@@ -108,11 +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,
@@ -162,6 +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" {
@@ -171,12 +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

@@ -105,11 +105,14 @@ type BenchmarkGPUResult struct {
 	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"`
+	Baseline               BenchmarkTelemetrySummary       `json:"baseline"`
-	Steady                 BenchmarkTelemetrySummary  `json:"steady"`
+	Steady                 BenchmarkTelemetrySummary       `json:"steady"`
-	Cooldown               BenchmarkTelemetrySummary  `json:"cooldown"`
+	PrecisionSteady        []BenchmarkPrecisionSteadyPhase `json:"precision_steady,omitempty"`
-	Throttle               BenchmarkThrottleCounters  `json:"throttle_counters"`
+	Cooldown               BenchmarkTelemetrySummary       `json:"cooldown"`
-	PrecisionResults       []BenchmarkPrecisionResult `json:"precision_results,omitempty"`
+	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"`
@@ -142,6 +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"`
@@ -152,19 +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"`
 }
 
@@ -182,6 +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/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,

audit/internal/webui/pages.go

@@ -1928,23 +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 {
@@ -2206,17 +2193,17 @@ benchmarkLoadGPUs();
 }
 
 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>`
 	}
@@ -2226,22 +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>`)
+	b.WriteString(`<table><thead><tr><th>Run</th><th>Time</th>`)
-	for _, col := range columns {
+	for i := 0; i <= maxGPUIndex; i++ {
-		b.WriteString(`<th>` + html.EscapeString(col.label) + `</th>`)
+		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 {
+		for idx := 0; idx <= maxGPUIndex; idx++ {
-			cell, ok := run.cells[col.key]
+			score, ok := run.gpuScores[idx]
-			if !ok || !cell.present {
+			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>`)
 	}
@@ -2249,22 +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) {
+func loadBenchmarkHistoryFromPaths(paths []string) (int, []benchmarkHistoryRun) {
-	columnByKey := make(map[string]benchmarkHistoryColumn)
 	runs := make([]benchmarkHistoryRun, 0, len(paths))
+	maxGPUIndex := -1
 	for _, path := range paths {
 		raw, err := os.ReadFile(path)
 		if err != nil {
@@ -2277,102 +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),
 		}
-
+		for _, gpu := range result.GPUs {
-		if result.ParallelGPUs {
+			run.gpuScores[gpu.Index] = gpu.Scores.CompositeScore
-			// All GPUs ran simultaneously — one column per server, score = avg composite.
+			if gpu.Index > maxGPUIndex {
-			gpuModelCount := make(map[string]int)
+				maxGPUIndex = gpu.Index
-			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}
 			}
 		}
 		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 ──────────────────────────────────────────────────────────────────────
 

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 0`,
-		`GPU #1 — NVIDIA H100 PCIe`,
+		`GPU 1`,
 		`#1`,
 		wantTime,
 		`1176.25`,

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