Split bee-bench into perf and power workflows

2026-04-14 17:33:13 +03:00
parent 54338dbae5
commit 95124d228f
17 changed files with 718 additions and 259 deletions
--- a/audit/internal/webui/api.go
+++ b/audit/internal/webui/api.go
@@ -110,7 +110,7 @@ func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {

 func shouldSplitHomogeneousNvidiaTarget(target string) bool {
 	switch strings.TrimSpace(target) {
-	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
+	case "nvidia", "nvidia-targeted-stress", "nvidia-bench-perf", "nvidia-bench-power", "nvidia-compute",
 		"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
 		"nvidia-bandwidth", "nvidia-stress":
 		return true
@@ -127,7 +127,7 @@ func defaultTaskPriority(target string, params taskParams) int {
 		return taskPriorityInstallToRAM
 	case "audit":
 		return taskPriorityAudit
-	case "nvidia-benchmark":
+	case "nvidia-bench-perf", "nvidia-bench-power":
 		return taskPriorityBenchmark
 	case "nvidia-stress", "amd-stress", "memory-stress", "sat-stress", "platform-stress", "nvidia-compute":
 		return taskPriorityBurn
@@ -573,131 +573,142 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 	}
 }

-func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
-	if h.opts.App == nil {
-		writeError(w, http.StatusServiceUnavailable, "app not configured")
-		return
-	}
-
-	var body struct {
-		Profile           string `json:"profile"`
-		SizeMB            int    `json:"size_mb"`
-		GPUIndices        []int  `json:"gpu_indices"`
-		ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
-		RunNCCL           *bool  `json:"run_nccl"`
-		ParallelGPUs      *bool  `json:"parallel_gpus"`
-		RampUp            *bool  `json:"ramp_up"`
-		DisplayName       string `json:"display_name"`
-	}
-	if r.Body != nil {
-		if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
-			writeError(w, http.StatusBadRequest, "invalid request body")
+func (h *handler) handleAPIBenchmarkNvidiaRunKind(target string) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		if h.opts.App == nil {
+			writeError(w, http.StatusServiceUnavailable, "app not configured")
 			return
 		}
-	}

-	runNCCL := true
-	if body.RunNCCL != nil {
-		runNCCL = *body.RunNCCL
-	}
-	parallelGPUs := false
-	if body.ParallelGPUs != nil {
-		parallelGPUs = *body.ParallelGPUs
-	}
-	rampUp := false
-	if body.RampUp != nil {
-		rampUp = *body.RampUp
-	}
-	// Build a descriptive base name that includes profile and mode so the task
-	// list is self-explanatory without opening individual task detail pages.
-	profile := strings.TrimSpace(body.Profile)
-	if profile == "" {
-		profile = "standard"
-	}
-	name := taskDisplayName("nvidia-benchmark", "", "")
-	if strings.TrimSpace(body.DisplayName) != "" {
-		name = body.DisplayName
-	}
-	// Append profile tag.
-	name = fmt.Sprintf("%s · %s", name, profile)
+		var body struct {
+			Profile           string `json:"profile"`
+			SizeMB            int    `json:"size_mb"`
+			GPUIndices        []int  `json:"gpu_indices"`
+			ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
+			RunNCCL           *bool  `json:"run_nccl"`
+			ParallelGPUs      *bool  `json:"parallel_gpus"`
+			RampUp            *bool  `json:"ramp_up"`
+			DisplayName       string `json:"display_name"`
+		}
+		if r.Body != nil {
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
+				writeError(w, http.StatusBadRequest, "invalid request body")
+				return
+			}
+		}

-	if rampUp && len(body.GPUIndices) > 1 {
-		// Ramp-up mode: resolve GPU list, then create one task per prefix
-		// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
-		gpus, err := apiListNvidiaGPUs(h.opts.App)
-		if err != nil {
-			writeError(w, http.StatusBadRequest, err.Error())
+		runNCCL := true
+		if body.RunNCCL != nil {
+			runNCCL = *body.RunNCCL
+		}
+		parallelGPUs := false
+		if body.ParallelGPUs != nil {
+			parallelGPUs = *body.ParallelGPUs
+		}
+		rampUp := false
+		if body.RampUp != nil {
+			rampUp = *body.RampUp
+		}
+		// Build a descriptive base name that includes profile and mode so the task
+		// list is self-explanatory without opening individual task detail pages.
+		profile := strings.TrimSpace(body.Profile)
+		if profile == "" {
+			profile = "standard"
+		}
+		name := taskDisplayName(target, "", "")
+		if strings.TrimSpace(body.DisplayName) != "" {
+			name = body.DisplayName
+		}
+		// Append profile tag.
+		name = fmt.Sprintf("%s · %s", name, profile)
+
+		if target == "nvidia-bench-power" && parallelGPUs {
+			writeError(w, http.StatusBadRequest, "power / thermal fit benchmark uses sequential or ramp-up modes only")
 			return
 		}
-		resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
-		if err != nil {
-			writeError(w, http.StatusBadRequest, err.Error())
-			return
-		}
-		if len(resolved) < 2 {
-			// Fall through to normal single-task path.
-			rampUp = false
-		} else {
-			now := time.Now()
-			rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
-			var allTasks []*Task
-			for step := 1; step <= len(resolved); step++ {
-				subset := resolved[:step]
-				stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
-				t := &Task{
-					ID:        newJobID("benchmark-nvidia"),
-					Name:      stepName,
-					Target:    "nvidia-benchmark",
-					Priority:  defaultTaskPriority("nvidia-benchmark", taskParams{}),
-					Status:    TaskPending,
-					CreatedAt: now,
-					params: taskParams{
-						GPUIndices:       append([]int(nil), subset...),
-						SizeMB:           body.SizeMB,
-						BenchmarkProfile: body.Profile,
-						RunNCCL:          runNCCL && step == len(resolved),
-						ParallelGPUs:     true,
-						RampStep:         step,
-						RampTotal:        len(resolved),
-						RampRunID:        rampRunID,
-						DisplayName:      stepName,
-					},
+
+		if rampUp && len(body.GPUIndices) > 1 {
+			// Ramp-up mode: resolve GPU list, then create one task per prefix
+			// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
+			gpus, err := apiListNvidiaGPUs(h.opts.App)
+			if err != nil {
+				writeError(w, http.StatusBadRequest, err.Error())
+				return
+			}
+			resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
+			if err != nil {
+				writeError(w, http.StatusBadRequest, err.Error())
+				return
+			}
+			if len(resolved) < 2 {
+				// Fall through to normal single-task path.
+				rampUp = false
+			} else {
+				now := time.Now()
+				rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
+				var allTasks []*Task
+				for step := 1; step <= len(resolved); step++ {
+					subset := resolved[:step]
+					stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
+					t := &Task{
+						ID:        newJobID("bee-bench-nvidia"),
+						Name:      stepName,
+						Target:    target,
+						Priority:  defaultTaskPriority(target, taskParams{}),
+						Status:    TaskPending,
+						CreatedAt: now,
+						params: taskParams{
+							GPUIndices:       append([]int(nil), subset...),
+							SizeMB:           body.SizeMB,
+							BenchmarkProfile: body.Profile,
+							RunNCCL:          runNCCL && step == len(resolved),
+							ParallelGPUs:     true,
+							RampStep:         step,
+							RampTotal:        len(resolved),
+							RampRunID:        rampRunID,
+							DisplayName:      stepName,
+						},
+					}
+					allTasks = append(allTasks, t)
 				}
-				allTasks = append(allTasks, t)
+				for _, t := range allTasks {
+					globalQueue.enqueue(t)
+				}
+				writeTaskRunResponse(w, allTasks)
+				return
 			}
-			for _, t := range allTasks {
-				globalQueue.enqueue(t)
-			}
-			writeTaskRunResponse(w, allTasks)
+		}
+
+		// For non-ramp tasks append mode tag.
+		if parallelGPUs {
+			name = fmt.Sprintf("%s · parallel", name)
+		} else {
+			name = fmt.Sprintf("%s · sequential", name)
+		}
+
+		params := taskParams{
+			GPUIndices:        body.GPUIndices,
+			ExcludeGPUIndices: body.ExcludeGPUIndices,
+			SizeMB:            body.SizeMB,
+			BenchmarkProfile:  body.Profile,
+			RunNCCL:           runNCCL,
+			ParallelGPUs:      parallelGPUs,
+			DisplayName:       body.DisplayName,
+		}
+		tasks, err := buildNvidiaTaskSet(target, defaultTaskPriority(target, params), time.Now(), params, name, h.opts.App, "bee-bench-nvidia")
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
 			return
 		}
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
 	}
+}

-	// For non-ramp tasks append mode tag.
-	if parallelGPUs {
-		name = fmt.Sprintf("%s · parallel", name)
-	} else {
-		name = fmt.Sprintf("%s · sequential", name)
-	}
-
-	params := taskParams{
-		GPUIndices:        body.GPUIndices,
-		ExcludeGPUIndices: body.ExcludeGPUIndices,
-		SizeMB:            body.SizeMB,
-		BenchmarkProfile:  body.Profile,
-		RunNCCL:           runNCCL,
-		ParallelGPUs:      parallelGPUs,
-		DisplayName:       body.DisplayName,
-	}
-	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", defaultTaskPriority("nvidia-benchmark", params), time.Now(), params, name, h.opts.App, "benchmark-nvidia")
-	if err != nil {
-		writeError(w, http.StatusBadRequest, err.Error())
-		return
-	}
-	for _, t := range tasks {
-		globalQueue.enqueue(t)
-	}
-	writeTaskRunResponse(w, tasks)
+func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
+	h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf").ServeHTTP(w, r)
 }

 func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
--- a/audit/internal/webui/api_test.go
+++ b/audit/internal/webui/api_test.go
@@ -64,7 +64,7 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	t.Cleanup(func() { apiListNvidiaGPUs = prevList })

 	h := &handler{opts: HandlerOptions{App: &app.App{}}}
-	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/perf/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
 	rec := httptest.NewRecorder()

 	h.handleAPIBenchmarkNvidiaRun(rec, req)
@@ -78,8 +78,8 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 		t.Fatalf("tasks=%d want 1", len(globalQueue.tasks))
 	}
 	task := globalQueue.tasks[0]
-	if task.Target != "nvidia-benchmark" {
-		t.Fatalf("target=%q want nvidia-benchmark", task.Target)
+	if task.Target != "nvidia-bench-perf" {
+		t.Fatalf("target=%q want nvidia-bench-perf", task.Target)
 	}
 	if got := task.params.GPUIndices; len(got) != 2 || got[0] != 1 || got[1] != 3 {
 		t.Fatalf("gpu indices=%v want [1 3]", got)
@@ -113,7 +113,7 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	t.Cleanup(func() { apiListNvidiaGPUs = prevList })

 	h := &handler{opts: HandlerOptions{App: &app.App{}}}
-	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/perf/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
 	rec := httptest.NewRecorder()

 	h.handleAPIBenchmarkNvidiaRun(rec, req)
@@ -147,6 +147,50 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	}
 }

+func TestHandleAPIBenchmarkPowerFitRampQueuesBenchmarkPowerFitTasks(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H100 PCIe"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/bee-bench/nvidia/power/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"ramp_up":true}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-power").ServeHTTP(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 3 {
+		t.Fatalf("tasks=%d want 3", len(globalQueue.tasks))
+	}
+	for i, task := range globalQueue.tasks {
+		if task.Target != "nvidia-bench-power" {
+			t.Fatalf("task[%d] target=%q", i, task.Target)
+		}
+		if task.Priority != taskPriorityBenchmark {
+			t.Fatalf("task[%d] priority=%d want %d", i, task.Priority, taskPriorityBenchmark)
+		}
+	}
+}
+
 func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
 	globalQueue.mu.Lock()
 	originalTasks := globalQueue.tasks
@@ -202,7 +246,8 @@ func TestDefaultTaskPriorityOrder(t *testing.T) {
 		defaultTaskPriority("cpu", taskParams{}),
 		defaultTaskPriority("cpu", taskParams{StressMode: true}),
 		defaultTaskPriority("nvidia-stress", taskParams{}),
-		defaultTaskPriority("nvidia-benchmark", taskParams{}),
+		defaultTaskPriority("nvidia-bench-perf", taskParams{}),
+		defaultTaskPriority("nvidia-bench-power", taskParams{}),
 	}
 	want := []int{
 		taskPriorityInstallToRAM,
@@ -211,13 +256,14 @@ func TestDefaultTaskPriorityOrder(t *testing.T) {
 		taskPriorityValidateStress,
 		taskPriorityBurn,
 		taskPriorityBenchmark,
+		taskPriorityBenchmark,
 	}
 	for i := range want {
 		if got[i] != want[i] {
 			t.Fatalf("priority[%d]=%d want %d", i, got[i], want[i])
 		}
 	}
-	if !(got[0] > got[1] && got[1] > got[2] && got[2] > got[3] && got[3] > got[4] && got[4] > got[5]) {
+	if !(got[0] > got[1] && got[1] > got[2] && got[2] > got[3] && got[3] > got[4] && got[4] > got[5] && got[5] == got[6]) {
 		t.Fatalf("priority order=%v", got)
 	}
 }
--- a/audit/internal/webui/kmsg_watcher.go
+++ b/audit/internal/webui/kmsg_watcher.go
@@ -232,7 +232,7 @@ func truncate(s string, max int) string {
 // isSATTarget returns true for task targets that run hardware acceptance tests.
 func isSATTarget(target string) bool {
 	switch target {
-	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
+	case "nvidia", "nvidia-targeted-stress", "nvidia-bench-perf", "nvidia-bench-power", "nvidia-compute", "nvidia-targeted-power", "nvidia-pulse",
 		"nvidia-interconnect", "nvidia-bandwidth", "nvidia-stress", "memory", "memory-stress", "storage",
 		"cpu", "sat-stress", "amd", "amd-mem", "amd-bandwidth", "amd-stress",
 		"platform-stress":
--- a/audit/internal/webui/pages.go
+++ b/audit/internal/webui/pages.go
@@ -1946,7 +1946,7 @@ func renderBenchmark(opts HandlerOptions) string {

 <div class="grid2">
  <div class="card">
-    <div class="card-head">NVIDIA Benchmark</div>
+    <div class="card-head">Benchmark Setup</div>
    <div class="card-body">
      <div class="form-row">
        <label>Profile</label>
@@ -1979,21 +1979,25 @@ func renderBenchmark(opts HandlerOptions) string {
        <span>Ramp-up — 1 GPU → 2 → … → all selected (separate tasks)</span>
      </label>
      <p id="benchmark-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 14px">Select one GPU for single-card benchmarking or several GPUs for a constrained multi-GPU run.</p>
-      <button id="benchmark-run-btn" class="btn btn-primary" onclick="runNvidiaBenchmark()" disabled>&#9654; Run Benchmark</button>
+      <div style="display:flex;gap:8px;flex-wrap:wrap;align-items:center">
+        <button id="benchmark-run-performance-btn" class="btn btn-primary" onclick="runNvidiaBenchmark('performance')" disabled>&#9654; Run Performance Benchmark</button>
+        <button id="benchmark-run-power-fit-btn" class="btn btn-secondary" onclick="runNvidiaBenchmark('power-fit')" disabled>&#9654; Run Power / Thermal Fit</button>
+      </div>
+      <span id="benchmark-run-nccl" hidden>nccl-auto</span>
      <span id="benchmark-run-status" style="margin-left:10px;font-size:12px;color:var(--muted)"></span>
    </div>
  </div>

  <div class="card">
-    <div class="card-head">Method</div>
+    <div class="card-head">Method Split</div>
    <div class="card-body">
-      <p style="font-size:13px;color:var(--muted);margin-bottom:10px">Each benchmark run performs warmup, sustained compute, telemetry capture, cooldown, and optional NCCL interconnect checks.</p>
+      <p style="font-size:13px;color:var(--muted);margin-bottom:10px">The benchmark page now exposes two fundamentally different test families so compute score and server power-fit are not mixed into one number.</p>
      <table>
-        <tr><th>Profile</th><th>Purpose</th></tr>
-        <tr><td>Standard</td><td>Fast, repeatable performance check for server-to-server comparison.</td></tr>
-        <tr><td>Stability</td><td>Longer run for thermal drift, power caps, and clock instability.</td></tr>
-        <tr><td>Overnight</td><td>Extended verification of long-run stability and late throttling.</td></tr>
+        <tr><th>Run Type</th><th>Engine</th><th>Question</th></tr>
+        <tr><td>Performance Benchmark</td><td><code>bee-gpu-burn</code></td><td>How much isolated compute performance does the GPU realize in this server?</td></tr>
+        <tr><td>Power / Thermal Fit</td><td><code>dcgmi targeted_power</code></td><td>How much power per GPU can this server sustain as GPU count ramps up?</td></tr>
      </table>
+      <p style="font-size:12px;color:var(--muted);margin-top:10px">Use ramp-up mode for capacity work: it creates 1 GPU → 2 GPU → … → all selected steps so analysis software can derive server total score and watts-per-GPU curves.</p>
    </div>
  </div>
 </div>
@@ -2036,21 +2040,24 @@ function benchmarkMode() {

 function benchmarkUpdateSelectionNote() {
  const selected = benchmarkSelectedGPUIndices();
-  const btn = document.getElementById('benchmark-run-btn');
+  const perfBtn = document.getElementById('benchmark-run-performance-btn');
+  const fitBtn = document.getElementById('benchmark-run-power-fit-btn');
  const note = document.getElementById('benchmark-selection-note');
  if (!selected.length) {
-    btn.disabled = true;
+    perfBtn.disabled = true;
+    fitBtn.disabled = true;
    note.textContent = 'Select at least one NVIDIA GPU to run the benchmark.';
    return;
  }
-  btn.disabled = false;
+  perfBtn.disabled = false;
+  fitBtn.disabled = false;
  const mode = benchmarkMode();
  if (mode === 'ramp-up') {
-    note.textContent = 'Ramp-up: ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). NCCL on final step.';
+    note.textContent = 'Ramp-up: ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). Performance uses compute benchmark; Power / Thermal Fit uses targeted_power per step.';
  } else if (mode === 'parallel') {
-    note.textContent = 'Parallel: all ' + selected.length + ' GPU(s) simultaneously.' + (selected.length > 1 ? ' NCCL included.' : '');
+    note.textContent = 'Parallel: all ' + selected.length + ' GPU(s) simultaneously. Only the performance benchmark supports this mode.';
  } else {
-    note.textContent = 'Sequential: each GPU benchmarked separately.' + (selected.length > 1 ? ' NCCL included on each.' : '');
+    note.textContent = 'Sequential: each selected GPU benchmarked separately.';
  }
 }

@@ -2124,7 +2131,7 @@ function benchmarkSelectNone() {
  benchmarkUpdateSelectionNote();
 }

-function runNvidiaBenchmark() {
+function runNvidiaBenchmark(kind) {
  const selected = benchmarkSelectedGPUIndices();
  const status = document.getElementById('benchmark-run-status');
  if (!selected.length) {
@@ -2134,21 +2141,26 @@ function runNvidiaBenchmark() {
  if (benchmarkES) { benchmarkES.close(); benchmarkES = null; }
  const mode = benchmarkMode();
  const rampUp = mode === 'ramp-up' && selected.length > 1;
-  const parallelGPUs = mode === 'parallel';
+  const parallelGPUs = mode === 'parallel' && kind === 'performance';
+  if (kind === 'power-fit' && mode === 'parallel') {
+    status.textContent = 'Power / Thermal Fit supports sequential or ramp-up only.';
+    return;
+  }
  const body = {
    profile: document.getElementById('benchmark-profile').value || 'standard',
    gpu_indices: selected,
-    run_nccl: selected.length > 1,
+    run_nccl: kind === 'performance' && selected.length > 1,
    parallel_gpus: parallelGPUs,
    ramp_up: rampUp,
-    display_name: 'NVIDIA Benchmark'
+    display_name: kind === 'power-fit' ? 'NVIDIA Power / Thermal Fit' : 'NVIDIA Performance Benchmark'
  };
  document.getElementById('benchmark-output').style.display = 'block';
-  document.getElementById('benchmark-title').textContent = '— ' + body.profile + ' [' + selected.join(', ') + ']';
+  document.getElementById('benchmark-title').textContent = '— ' + body.display_name + ' · ' + body.profile + ' [' + selected.join(', ') + ']';
  const term = document.getElementById('benchmark-terminal');
-  term.textContent = 'Enqueuing benchmark for GPUs ' + selected.join(', ') + '...\n';
+  term.textContent = 'Enqueuing ' + body.display_name + ' for GPUs ' + selected.join(', ') + '...\n';
  status.textContent = 'Queueing...';
-  fetch('/api/benchmark/nvidia/run', {
+  const endpoint = kind === 'power-fit' ? '/api/bee-bench/nvidia/power/run' : '/api/bee-bench/nvidia/perf/run';
+  fetch(endpoint, {
    method: 'POST',
    headers: {'Content-Type':'application/json'},
    body: JSON.stringify(body)
@@ -2202,7 +2214,7 @@ benchmarkLoadGPUs();
 func renderBenchmarkResultsCard(exportDir string) string {
 	maxIdx, runs := loadBenchmarkHistory(exportDir)
 	return renderBenchmarkResultsCardFromRuns(
-		"Benchmark Results",
+		"Perf Results",
 		"Composite score by saved benchmark run and GPU.",
 		"No saved benchmark runs yet.",
 		maxIdx,
@@ -2244,11 +2256,11 @@ func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string,
 }

 func loadBenchmarkHistory(exportDir string) (int, []benchmarkHistoryRun) {
-	baseDir := app.DefaultBenchmarkBaseDir
+	baseDir := app.DefaultBeeBenchPerfDir
 	if strings.TrimSpace(exportDir) != "" {
-		baseDir = filepath.Join(exportDir, "bee-benchmark")
+		baseDir = filepath.Join(exportDir, "bee-bench", "perf")
 	}
-	paths, err := filepath.Glob(filepath.Join(baseDir, "gpu-benchmark-*", "result.json"))
+	paths, err := filepath.Glob(filepath.Join(baseDir, "perf-*", "result.json"))
 	if err != nil || len(paths) == 0 {
 		return -1, nil
 	}
--- a/audit/internal/webui/server.go
+++ b/audit/internal/webui/server.go
@@ -261,7 +261,8 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/sat/platform-stress/run", h.handleAPISATRun("platform-stress"))
 	mux.HandleFunc("GET /api/sat/stream", h.handleAPISATStream)
 	mux.HandleFunc("POST /api/sat/abort", h.handleAPISATAbort)
-	mux.HandleFunc("POST /api/benchmark/nvidia/run", h.handleAPIBenchmarkNvidiaRun)
+	mux.HandleFunc("POST /api/bee-bench/nvidia/perf/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf"))
+	mux.HandleFunc("POST /api/bee-bench/nvidia/power/run", h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-power"))

 	// Tasks
 	mux.HandleFunc("GET /api/tasks", h.handleAPITasksList)
--- a/audit/internal/webui/server_test.go
+++ b/audit/internal/webui/server_test.go
@@ -648,8 +648,11 @@ func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
 		`href="/benchmark"`,
 		`id="benchmark-gpu-list"`,
 		`/api/gpu/nvidia`,
-		`/api/benchmark/nvidia/run`,
+		`/api/bee-bench/nvidia/perf/run`,
+		`/api/bee-bench/nvidia/power/run`,
 		`benchmark-run-nccl`,
+		`Run Performance Benchmark`,
+		`Run Power / Thermal Fit`,
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("benchmark page missing %q: %s", needle, body)
@@ -660,7 +663,7 @@ func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
 func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	dir := t.TempDir()
 	exportDir := filepath.Join(dir, "export")
-	runDir := filepath.Join(exportDir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	runDir := filepath.Join(exportDir, "bee-bench", "perf", "perf-20260406-120000")
 	if err := os.MkdirAll(runDir, 0755); err != nil {
 		t.Fatal(err)
 	}
@@ -702,7 +705,7 @@ func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
 	body := rec.Body.String()
 	wantTime := result.GeneratedAt.Local().Format("2006-01-02 15:04:05")
 	for _, needle := range []string{
-		`Benchmark Results`,
+		`Perf Results`,
 		`Composite score by saved benchmark run and GPU.`,
 		`GPU 0`,
 		`GPU 1`,
--- a/audit/internal/webui/task_report.go
+++ b/audit/internal/webui/task_report.go
@@ -251,7 +251,9 @@ func renderTaskReportFragment(report taskReport, charts map[string]string, logTe
 }

 func renderTaskBenchmarkResultsCard(target, logText string) string {
-	if strings.TrimSpace(target) != "nvidia-benchmark" {
+	switch strings.TrimSpace(target) {
+	case "nvidia-bench-perf":
+	default:
 		return ""
 	}
 	resultPath := taskBenchmarkResultPath(logText)
@@ -263,7 +265,7 @@ func renderTaskBenchmarkResultsCard(target, logText string) string {
 		return ""
 	}
 	return renderBenchmarkResultsCardFromRuns(
-		"Benchmark Results",
+		"Perf Results",
 		"Composite score for this benchmark task.",
 		"No benchmark results were saved for this task.",
 		columns,
--- a/audit/internal/webui/tasks.go
+++ b/audit/internal/webui/tasks.go
@@ -32,7 +32,8 @@ const (
 var taskNames = map[string]string{
 	"nvidia":                 "NVIDIA SAT",
 	"nvidia-targeted-stress": "NVIDIA Targeted Stress Validate (dcgmi diag targeted_stress)",
-	"nvidia-benchmark":       "NVIDIA Benchmark",
+	"nvidia-bench-perf":      "NVIDIA Bee Bench Perf",
+	"nvidia-bench-power":     "NVIDIA Bee Bench Power",
 	"nvidia-compute":         "NVIDIA Max Compute Load (dcgmproftester)",
 	"nvidia-targeted-power":  "NVIDIA Targeted Power (dcgmi diag targeted_power)",
 	"nvidia-pulse":           "NVIDIA Pulse Test (dcgmi diag pulse_test)",
@@ -628,7 +629,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			dur = 300
 		}
 		archive, err = a.RunNvidiaTargetedStressValidatePack(ctx, "", dur, t.params.GPUIndices, j.append)
-	case "nvidia-benchmark":
+	case "nvidia-bench-perf":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
@@ -644,6 +645,31 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			RampTotal:         t.params.RampTotal,
 			RampRunID:         t.params.RampRunID,
 		}, j.append)
+	case "nvidia-bench-power":
+		if a == nil {
+			err = fmt.Errorf("app not configured")
+			break
+		}
+		dur := t.params.Duration
+		if dur <= 0 {
+			switch strings.TrimSpace(strings.ToLower(t.params.BenchmarkProfile)) {
+			case platform.NvidiaBenchmarkProfileStability:
+				dur = 300
+			case platform.NvidiaBenchmarkProfileOvernight:
+				dur = 600
+			default:
+				dur = 120
+			}
+		}
+		rampPlan, planErr := resolveNvidiaRampPlan(t.params.BenchmarkProfile, t.params.RampTotal > 0, t.params.GPUIndices)
+		if planErr != nil {
+			err = planErr
+			break
+		}
+		if t.params.RampTotal > 0 && t.params.RampStep > 0 && dur <= 0 {
+			dur = rampPlan.DurationSec
+		}
+		archive, err = a.RunNvidiaTargetedPowerPack(ctx, app.DefaultBeeBenchPowerDir, dur, t.params.GPUIndices, j.append)
 	case "nvidia-compute":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
--- a/audit/internal/webui/tasks_test.go
+++ b/audit/internal/webui/tasks_test.go
@@ -366,7 +366,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	taskReportMetricsDBPath = metricsPath
 	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })

-	benchmarkDir := filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	benchmarkDir := filepath.Join(dir, "bee-bench", "perf", "perf-20260406-120000")
 	if err := os.MkdirAll(benchmarkDir, 0755); err != nil {
 		t.Fatal(err)
 	}
@@ -398,14 +398,14 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	}
 	task := &Task{
 		ID:           "task-bench",
-		Name:         "NVIDIA Benchmark",
-		Target:       "nvidia-benchmark",
+		Name:         "NVIDIA Bee Bench Perf",
+		Target:       "nvidia-bench-perf",
 		Status:       TaskDone,
 		CreatedAt:    time.Now().UTC().Add(-time.Minute),
 		ArtifactsDir: artifactsDir,
 	}
 	ensureTaskReportPaths(task)
-	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000.tar.gz") + "\n"
+	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-bench", "perf", "perf-20260406-120000.tar.gz") + "\n"
 	if err := os.WriteFile(task.LogPath, []byte(logText), 0644); err != nil {
 		t.Fatal(err)
 	}
@@ -420,7 +420,7 @@ func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
 	}
 	html := string(body)
 	for _, needle := range []string{
-		`Benchmark Results`,
+		`Perf Results`,
 		`Composite score for this benchmark task.`,
 		`GPU 0`,
 		`1176.25`,