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>
This commit is contained in:
2026-04-13 10:49:49 +03:00
parent 02e44b1172
commit bf6ecab4f0
9 changed files with 390 additions and 144 deletions

View File

@@ -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 {
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
}
}
// MixedEfficiency = MixedScore / SyntheticScore.
// Measures how well the GPU sustains throughput under concurrent mixed load.
// A healthy GPU scores ~0.80.95; 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