Normalize task IDs and artifact folder prefixes

audit/internal/platform/benchmark.go

@@ -679,7 +679,10 @@ func runBenchmarkInterconnect(ctx context.Context, verboseLog, runDir string, gp
 		"-g", strconv.Itoa(len(gpuIndices)),
 		"--iters", strconv.Itoa(maxInt(20, spec.NCCLSec/10)),
 	}
-	env := []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	env := []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 	logFunc(fmt.Sprintf("NCCL interconnect: gpus=%s", joinIndexList(gpuIndices)))
 	out, err := runSATCommandCtx(ctx, verboseLog, "nccl-all-reduce.log", cmd, env, logFunc)
 	_ = os.WriteFile(filepath.Join(runDir, "nccl-all-reduce.log"), out, 0644)

audit/internal/platform/nvidia_stress.go

@@ -16,12 +16,12 @@ func (s *System) RunNvidiaStressPack(ctx context.Context, baseDir string, opts N
 		return "", err
 	}
 
-	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
+	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
 		job,
-		{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func nvidiaStressArchivePrefix(loader string) string {

audit/internal/platform/sat.go

@@ -278,13 +278,13 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 	if gpuCount < 1 {
 		gpuCount = 1
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-all-reduce-perf.log", cmd: []string{
+	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-all-reduce-perf.log", cmd: []string{
 			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
 			"-g", strconv.Itoa(gpuCount), "--iters", "20",
 		}},
-	}, logFunc)
+	), logFunc)
 }
 
 func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -296,18 +296,18 @@ func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir strin
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
+		satJob{
 			name:       "03-dcgmproftester.log",
 			cmd:        profCmd,
 			env:        nvidiaVisibleDevicesEnv(selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -315,16 +315,16 @@ func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string,
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-targeted-power.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("targeted_power", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -332,16 +332,16 @@ func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, dur
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-pulse-test.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("pulse_test", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -349,16 +349,16 @@ func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpu
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-nvbandwidth.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("nvbandwidth", 0, selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
@@ -389,16 +389,16 @@ func (s *System) RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDi
 			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
 		}
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-stress", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-stress", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-targeted-stress.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("targeted_stress", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func resolveDCGMGPUIndices(gpuIndices []int) ([]int, error) {
@@ -568,14 +568,24 @@ type satStats struct {
 	Unsupported int
 }
 
+func withNvidiaPersistenceMode(jobs ...satJob) []satJob {
+	out := make([]satJob, 0, len(jobs)+1)
+	out = append(out, satJob{
+		name: "00-nvidia-smi-persistence-mode.log",
+		cmd:  []string{"nvidia-smi", "-pm", "1"},
+	})
+	out = append(out, jobs...)
+	return out
+}
+
 func nvidiaSATJobs() []satJob {
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
-		{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
-	}
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
+		satJob{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
+	)
 }
 
 func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
@@ -590,12 +600,12 @@ func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
 		}
 		diagArgs = append(diagArgs, "-i", strings.Join(ids, ","))
 	}
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-dcgmi-diag.log", cmd: diagArgs},
-	}
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-dcgmi-diag.log", cmd: diagArgs},
+	)
 }
 
 func nvidiaDCGMNamedDiagCommand(name string, durationSec int, gpuIndices []int) []string {
@@ -620,7 +630,10 @@ func nvidiaVisibleDevicesEnv(gpuIndices []int) []string {
 	if len(gpuIndices) == 0 {
 		return nil
 	}
-	return []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	return []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 }
 
 func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []satJob, logFunc func(string)) (string, error) {
@@ -661,6 +674,9 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		if writeErr := os.WriteFile(filepath.Join(runDir, job.name), out, 0644); writeErr != nil {
 			return "", writeErr
 		}
+		if ctx.Err() != nil {
+			return "", ctx.Err()
+		}
 		status, rc := classifySATResult(job.name, out, err)
 		stats.Add(status)
 		key := strings.TrimSuffix(strings.TrimPrefix(job.name, "0"), ".log")

audit/internal/platform/sat_test.go

@@ -1,12 +1,14 @@
 package platform
 
 import (
+	"context"
 	"errors"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestStorageSATCommands(t *testing.T) {
@@ -28,13 +30,19 @@ func TestRunNvidiaAcceptancePackIncludesGPUStress(t *testing.T) {
 
 	jobs := nvidiaSATJobs()
 
-	if len(jobs) != 5 {
-		t.Fatalf("jobs=%d want 5", len(jobs))
+	if len(jobs) != 6 {
+		t.Fatalf("jobs=%d want 6", len(jobs))
 	}
-	if got := jobs[4].cmd[0]; got != "bee-gpu-burn" {
+	if got := jobs[0].cmd[0]; got != "nvidia-smi" {
+		t.Fatalf("preflight command=%q want nvidia-smi", got)
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := jobs[5].cmd[0]; got != "bee-gpu-burn" {
 		t.Fatalf("gpu stress command=%q want bee-gpu-burn", got)
 	}
-	if got := jobs[3].cmd[1]; got != "--output-file" {
+	if got := jobs[4].cmd[1]; got != "--output-file" {
 		t.Fatalf("bug report flag=%q want --output-file", got)
 	}
 }
@@ -82,7 +90,7 @@ func TestAMDStressJobsIncludeBandwidthAndGST(t *testing.T) {
 
 func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	jobs := nvidiaSATJobs()
-	got := jobs[4].cmd
+	got := jobs[5].cmd
 	want := []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}
 	if len(got) != len(want) {
 		t.Fatalf("cmd len=%d want %d", len(got), len(want))
@@ -94,6 +102,19 @@ func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	}
 }
 
+func TestNvidiaDCGMJobsEnablePersistenceModeBeforeDiag(t *testing.T) {
+	jobs := nvidiaDCGMJobs(3, []int{2, 0})
+	if len(jobs) != 5 {
+		t.Fatalf("jobs=%d want 5", len(jobs))
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := strings.Join(jobs[4].cmd, " "); got != "dcgmi diag -r 3 -i 2,0" {
+		t.Fatalf("diag=%q want %q", got, "dcgmi diag -r 3 -i 2,0")
+	}
+}
+
 func TestBuildNvidiaStressJobUsesSelectedLoaderAndDevices(t *testing.T) {
 	t.Parallel()
 
@@ -234,11 +255,14 @@ func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
 
 func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
 	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
-	if len(env) != 1 {
-		t.Fatalf("env len=%d want 1 (%v)", len(env), env)
+	if len(env) != 2 {
+		t.Fatalf("env len=%d want 2 (%v)", len(env), env)
 	}
-	if env[0] != "CUDA_VISIBLE_DEVICES=0,2,4" {
-		t.Fatalf("env[0]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[0])
+	if env[0] != "CUDA_DEVICE_ORDER=PCI_BUS_ID" {
+		t.Fatalf("env[0]=%q want CUDA_DEVICE_ORDER=PCI_BUS_ID", env[0])
+	}
+	if env[1] != "CUDA_VISIBLE_DEVICES=0,2,4" {
+		t.Fatalf("env[1]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[1])
 	}
 }
 
@@ -331,6 +355,38 @@ func TestClassifySATResult(t *testing.T) {
 	}
 }
 
+func TestRunAcceptancePackCtxReturnsContextErrorWithoutArchive(t *testing.T) {
+	dir := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+
+	done := make(chan struct{})
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		cancel()
+		close(done)
+	}()
+
+	archive, err := runAcceptancePackCtx(ctx, dir, "cancelled-pack", []satJob{
+		{name: "01-sleep.log", cmd: []string{"sh", "-c", "sleep 5"}},
+	}, nil)
+	<-done
+
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err=%v want context.Canceled", err)
+	}
+	if archive != "" {
+		t.Fatalf("archive=%q want empty", archive)
+	}
+	matches, globErr := filepath.Glob(filepath.Join(dir, "cancelled-pack-*.tar.gz"))
+	if globErr != nil {
+		t.Fatalf("Glob error: %v", globErr)
+	}
+	if len(matches) != 0 {
+		t.Fatalf("archives=%v want none", matches)
+	}
+}
+
 func TestParseStorageDevicesSkipsUSBDisks(t *testing.T) {
 	t.Parallel()
 

audit/internal/webui/api.go

@@ -11,6 +11,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"regexp"
+	"sort"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -21,13 +22,238 @@ import (
 )
 
 var ansiEscapeRE = regexp.MustCompile(`\x1b\[[0-9;]*[a-zA-Z]|\x1b[()][A-Z0-9]|\x1b[DABC]`)
+var apiListNvidiaGPUs = func(a *app.App) ([]platform.NvidiaGPU, error) {
+	if a == nil {
+		return nil, fmt.Errorf("app not configured")
+	}
+	return a.ListNvidiaGPUs()
+}
 
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
 
-func newJobID(prefix string) string {
-	return fmt.Sprintf("%s-%d", prefix, jobCounter.Add(1))
+func newJobID(_ string) string {
+	start := int((jobCounter.Add(1) - 1) % 1000)
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	for offset := 0; offset < 1000; offset++ {
+		n := (start + offset) % 1000
+		id := fmt.Sprintf("TASK-%03d", n)
+		if !taskIDInUseLocked(id) {
+			return id
+		}
+	}
+	return fmt.Sprintf("TASK-%03d", start)
+}
+
+func taskIDInUseLocked(id string) bool {
+	for _, t := range globalQueue.tasks {
+		if t != nil && t.ID == id {
+			return true
+		}
+	}
+	return false
+}
+
+type taskRunResponse struct {
+	TaskID    string   `json:"task_id,omitempty"`
+	JobID     string   `json:"job_id,omitempty"`
+	TaskIDs   []string `json:"task_ids,omitempty"`
+	JobIDs    []string `json:"job_ids,omitempty"`
+	TaskCount int      `json:"task_count,omitempty"`
+}
+
+type nvidiaTaskSelection struct {
+	GPUIndices []int
+	Label      string
+}
+
+func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {
+	if len(tasks) == 0 {
+		writeJSON(w, taskRunResponse{})
+		return
+	}
+	ids := make([]string, 0, len(tasks))
+	for _, t := range tasks {
+		if t == nil || strings.TrimSpace(t.ID) == "" {
+			continue
+		}
+		ids = append(ids, t.ID)
+	}
+	resp := taskRunResponse{TaskCount: len(ids)}
+	if len(ids) > 0 {
+		resp.TaskID = ids[0]
+		resp.JobID = ids[0]
+		resp.TaskIDs = ids
+		resp.JobIDs = ids
+	}
+	writeJSON(w, resp)
+}
+
+func shouldSplitHomogeneousNvidiaTarget(target string) bool {
+	switch strings.TrimSpace(target) {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
+		"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
+		"nvidia-bandwidth", "nvidia-stress":
+		return true
+	default:
+		return false
+	}
+}
+
+func expandHomogeneousNvidiaSelections(gpus []platform.NvidiaGPU, include, exclude []int) ([]nvidiaTaskSelection, error) {
+	if len(gpus) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs detected")
+	}
+	indexed := make(map[int]platform.NvidiaGPU, len(gpus))
+	allIndices := make([]int, 0, len(gpus))
+	for _, gpu := range gpus {
+		indexed[gpu.Index] = gpu
+		allIndices = append(allIndices, gpu.Index)
+	}
+	sort.Ints(allIndices)
+
+	selected := allIndices
+	if len(include) > 0 {
+		selected = make([]int, 0, len(include))
+		seen := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			if _, ok := indexed[idx]; !ok {
+				continue
+			}
+			if _, dup := seen[idx]; dup {
+				continue
+			}
+			seen[idx] = struct{}{}
+			selected = append(selected, idx)
+		}
+		sort.Ints(selected)
+	}
+	if len(exclude) > 0 {
+		skip := make(map[int]struct{}, len(exclude))
+		for _, idx := range exclude {
+			skip[idx] = struct{}{}
+		}
+		filtered := selected[:0]
+		for _, idx := range selected {
+			if _, ok := skip[idx]; ok {
+				continue
+			}
+			filtered = append(filtered, idx)
+		}
+		selected = filtered
+	}
+	if len(selected) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs selected")
+	}
+
+	modelGroups := make(map[string][]platform.NvidiaGPU)
+	modelOrder := make([]string, 0)
+	for _, idx := range selected {
+		gpu := indexed[idx]
+		model := strings.TrimSpace(gpu.Name)
+		if model == "" {
+			model = fmt.Sprintf("GPU %d", gpu.Index)
+		}
+		if _, ok := modelGroups[model]; !ok {
+			modelOrder = append(modelOrder, model)
+		}
+		modelGroups[model] = append(modelGroups[model], gpu)
+	}
+	sort.Slice(modelOrder, func(i, j int) bool {
+		left := modelGroups[modelOrder[i]]
+		right := modelGroups[modelOrder[j]]
+		if len(left) == 0 || len(right) == 0 {
+			return modelOrder[i] < modelOrder[j]
+		}
+		return left[0].Index < right[0].Index
+	})
+
+	var groups []nvidiaTaskSelection
+	var singles []nvidiaTaskSelection
+	for _, model := range modelOrder {
+		group := modelGroups[model]
+		sort.Slice(group, func(i, j int) bool { return group[i].Index < group[j].Index })
+		indices := make([]int, 0, len(group))
+		for _, gpu := range group {
+			indices = append(indices, gpu.Index)
+		}
+		if len(indices) >= 2 {
+			groups = append(groups, nvidiaTaskSelection{
+				GPUIndices: indices,
+				Label:      fmt.Sprintf("%s; GPUs %s", model, joinTaskIndices(indices)),
+			})
+			continue
+		}
+		gpu := group[0]
+		singles = append(singles, nvidiaTaskSelection{
+			GPUIndices: []int{gpu.Index},
+			Label:      fmt.Sprintf("GPU %d — %s", gpu.Index, model),
+		})
+	}
+	return append(groups, singles...), nil
+}
+
+func joinTaskIndices(indices []int) string {
+	parts := make([]string, 0, len(indices))
+	for _, idx := range indices {
+		parts = append(parts, fmt.Sprintf("%d", idx))
+	}
+	return strings.Join(parts, ",")
+}
+
+func formatSplitTaskName(baseName, selectionLabel string) string {
+	baseName = strings.TrimSpace(baseName)
+	selectionLabel = strings.TrimSpace(selectionLabel)
+	if baseName == "" {
+		return selectionLabel
+	}
+	if selectionLabel == "" {
+		return baseName
+	}
+	return baseName + " (" + selectionLabel + ")"
+}
+
+func buildNvidiaTaskSet(target string, priority int, createdAt time.Time, params taskParams, baseName string, appRef *app.App, idPrefix string) ([]*Task, error) {
+	if !shouldSplitHomogeneousNvidiaTarget(target) {
+		t := &Task{
+			ID:        newJobID(idPrefix),
+			Name:      baseName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    params,
+		}
+		return []*Task{t}, nil
+	}
+	gpus, err := apiListNvidiaGPUs(appRef)
+	if err != nil {
+		return nil, err
+	}
+	selections, err := expandHomogeneousNvidiaSelections(gpus, params.GPUIndices, params.ExcludeGPUIndices)
+	if err != nil {
+		return nil, err
+	}
+	tasks := make([]*Task, 0, len(selections))
+	for _, selection := range selections {
+		taskParamsCopy := params
+		taskParamsCopy.GPUIndices = append([]int(nil), selection.GPUIndices...)
+		taskParamsCopy.ExcludeGPUIndices = nil
+		displayName := formatSplitTaskName(baseName, selection.Label)
+		taskParamsCopy.DisplayName = displayName
+		tasks = append(tasks, &Task{
+			ID:        newJobID(idPrefix),
+			Name:      displayName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    taskParamsCopy,
+		})
+	}
+	return tasks, nil
 }
 
 // ── SSE helpers ───────────────────────────────────────────────────────────────
@@ -207,28 +433,28 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		}
 
 		name := taskDisplayName(target, body.Profile, body.Loader)
-		t := &Task{
-			ID:        newJobID("sat-" + target),
-			Name:      name,
-			Target:    target,
-			Status:    TaskPending,
-			CreatedAt: time.Now(),
-			params: taskParams{
-				Duration:           body.Duration,
-				DiagLevel:          body.DiagLevel,
-				GPUIndices:         body.GPUIndices,
-				ExcludeGPUIndices:  body.ExcludeGPUIndices,
-				Loader:             body.Loader,
-				BurnProfile:        body.Profile,
-				DisplayName:        body.DisplayName,
-				PlatformComponents: body.PlatformComponents,
-			},
-		}
 		if strings.TrimSpace(body.DisplayName) != "" {
-			t.Name = body.DisplayName
+			name = body.DisplayName
 		}
-		globalQueue.enqueue(t)
-		writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
+		params := taskParams{
+			Duration:           body.Duration,
+			DiagLevel:          body.DiagLevel,
+			GPUIndices:         body.GPUIndices,
+			ExcludeGPUIndices:  body.ExcludeGPUIndices,
+			Loader:             body.Loader,
+			BurnProfile:        body.Profile,
+			DisplayName:        body.DisplayName,
+			PlatformComponents: body.PlatformComponents,
+		}
+		tasks, err := buildNvidiaTaskSet(target, 0, time.Now(), params, name, h.opts.App, "sat-"+target)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
 	}
 }
 
@@ -257,27 +483,26 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 	if body.RunNCCL != nil {
 		runNCCL = *body.RunNCCL
 	}
-	t := &Task{
-		ID:        newJobID("benchmark-nvidia"),
-		Name:      taskDisplayName("nvidia-benchmark", "", ""),
-		Target:    "nvidia-benchmark",
-		Priority:  15,
-		Status:    TaskPending,
-		CreatedAt: time.Now(),
-		params: taskParams{
-			GPUIndices:        body.GPUIndices,
-			ExcludeGPUIndices: body.ExcludeGPUIndices,
-			SizeMB:            body.SizeMB,
-			BenchmarkProfile:  body.Profile,
-			RunNCCL:           runNCCL,
-			DisplayName:       body.DisplayName,
-		},
-	}
+	name := taskDisplayName("nvidia-benchmark", "", "")
 	if strings.TrimSpace(body.DisplayName) != "" {
-		t.Name = body.DisplayName
+		name = body.DisplayName
 	}
-	globalQueue.enqueue(t)
-	writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
+	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
+		GPUIndices:        body.GPUIndices,
+		ExcludeGPUIndices: body.ExcludeGPUIndices,
+		SizeMB:            body.SizeMB,
+		BenchmarkProfile:  body.Profile,
+		RunNCCL:           runNCCL,
+		DisplayName:       body.DisplayName,
+	}, 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) handleAPISATStream(w http.ResponseWriter, r *http.Request) {

audit/internal/webui/api_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http/httptest"
 	"strings"
 	"testing"
@@ -74,6 +75,14 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 		globalQueue.tasks = originalTasks
 		globalQueue.mu.Unlock()
 	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 3, Name: "NVIDIA H100 PCIe"},
+		}, nil
+	}
+	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}`))
@@ -101,6 +110,97 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	}
 }
 
+func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(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 H200 NVL"},
+		}, nil
+	}
+	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}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRun(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	var resp taskRunResponse
+	if err := json.Unmarshal(rec.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(resp.TaskIDs) != 2 {
+		t.Fatalf("task_ids=%v want 2 items", resp.TaskIDs)
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
+
+func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(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 H200 NVL"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/sat/nvidia-targeted-power/run", strings.NewReader(`{"profile":"acceptance","gpu_indices":[0,1,2]}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPISATRun("nvidia-targeted-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) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
+
 func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {
 	h := &handler{}
 	h.pushFanRings([]platform.FanReading{

audit/internal/webui/pages.go

@@ -8,9 +8,12 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"strconv"
 	"strings"
+	"time"
 
 	"bee/audit/internal/app"
+	"bee/audit/internal/platform"
 	"bee/audit/internal/schema"
 )
 
@@ -161,7 +164,7 @@ func renderPage(page string, opts HandlerOptions) string {
 	case "benchmark":
 		pageID = "benchmark"
 		title = "Benchmark"
-		body = renderBenchmark()
+		body = renderBenchmark(opts)
 	case "tasks":
 		pageID = "tasks"
 		title = "Tasks"
@@ -1072,14 +1075,14 @@ func renderValidate(opts HandlerOptions) string {
 		inv.NVIDIA,
 		`Select which NVIDIA GPUs to include in Validate. The same selection is used by both NVIDIA GPU cards below and by Validate one by one.`,
 		`<code>nvidia-smi --query-gpu=index,name,memory.total</code>`,
-	`<div id="sat-gpu-list"><p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs…</p></div><div style="display:flex;gap:8px;flex-wrap:wrap;margin-top:8px"><button type="button" class="btn btn-sm btn-secondary" onclick="satSelectAllGPUs()">Select all</button><button type="button" class="btn btn-sm btn-secondary" onclick="satSelectNoGPUs()">Clear</button></div><div id="sat-gpu-selection-note" style="font-size:12px;color:var(--muted);margin-top:8px"></div>`,
-	)) +
-	renderSATCard("nvidia", "NVIDIA GPU", "runNvidiaValidateSet('nvidia')", "", renderValidateCardBody(
-		inv.NVIDIA,
-		`Runs NVIDIA diagnostics and board inventory checks.`,
-		`<code>nvidia-smi</code>, <code>dmidecode</code>, <code>dcgmi diag</code>`,
-		`Runs one GPU at a time on the selected NVIDIA GPUs. Diag level is taken from Validate Profile.`,
+		`<div id="sat-gpu-list"><p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs…</p></div><div style="display:flex;gap:8px;flex-wrap:wrap;margin-top:8px"><button type="button" class="btn btn-sm btn-secondary" onclick="satSelectAllGPUs()">Select all</button><button type="button" class="btn btn-sm btn-secondary" onclick="satSelectNoGPUs()">Clear</button></div><div id="sat-gpu-selection-note" style="font-size:12px;color:var(--muted);margin-top:8px"></div>`,
 	)) +
+		renderSATCard("nvidia", "NVIDIA GPU", "runNvidiaValidateSet('nvidia')", "", renderValidateCardBody(
+			inv.NVIDIA,
+			`Runs NVIDIA diagnostics and board inventory checks.`,
+			`<code>nvidia-smi</code>, <code>dmidecode</code>, <code>dcgmi diag</code>`,
+			`Runs one GPU at a time on the selected NVIDIA GPUs. Diag level is taken from Validate Profile.`,
+		)) +
 		renderSATCard("nvidia-targeted-stress", "NVIDIA GPU Targeted Stress", "runNvidiaValidateSet('nvidia-targeted-stress')", "", renderValidateCardBody(
 			inv.NVIDIA,
 			`Runs a controlled NVIDIA DCGM load in Validate to check stability under moderate stress.`,
@@ -1569,7 +1572,25 @@ func renderSATCard(id, label, runAction, headerActions, body string) string {
 
 // ── Benchmark ─────────────────────────────────────────────────────────────────
 
-func renderBenchmark() string {
+type benchmarkHistoryColumn struct {
+	key   string
+	label string
+	name  string
+	index int
+}
+
+type benchmarkHistoryCell struct {
+	score   float64
+	present bool
+}
+
+type benchmarkHistoryRun struct {
+	generatedAt time.Time
+	displayTime string
+	cells       map[string]benchmarkHistoryCell
+}
+
+func renderBenchmark(opts HandlerOptions) string {
 	return `<p style="color:var(--muted);font-size:13px;margin-bottom:16px">Benchmark runs generate a human-readable TXT report and machine-readable result bundle. Tasks continue in the background — view progress in <a href="/tasks">Tasks</a>.</p>
 
 <div class="grid2">
@@ -1618,6 +1639,8 @@ func renderBenchmark() string {
   </div>
 </div>
 
+` + renderBenchmarkResultsCard(opts.ExportDir) + `
+
 <div id="benchmark-output" style="display:none;margin-top:16px" class="card">
   <div class="card-head">Benchmark Output <span id="benchmark-title"></span></div>
   <div class="card-body"><div id="benchmark-terminal" class="terminal"></div></div>
@@ -1633,6 +1656,12 @@ func renderBenchmark() string {
 <script>
 let benchmarkES = null;
 
+function benchmarkTaskIDs(payload) {
+  if (payload && Array.isArray(payload.task_ids) && payload.task_ids.length) return payload.task_ids;
+  if (payload && payload.task_id) return [payload.task_id];
+  return [];
+}
+
 function benchmarkSelectedGPUIndices() {
   return Array.from(document.querySelectorAll('.benchmark-gpu-checkbox'))
     .filter(function(el) { return el.checked && !el.disabled; })
@@ -1732,17 +1761,37 @@ function runNvidiaBenchmark() {
       return payload;
     });
   }).then(function(d) {
-    status.textContent = 'Task ' + d.task_id + ' queued.';
-    term.textContent += 'Task ' + d.task_id + ' queued. Streaming log...\n';
-    benchmarkES = new EventSource('/api/tasks/' + d.task_id + '/stream');
-    benchmarkES.onmessage = function(e) { term.textContent += e.data + '\n'; term.scrollTop = term.scrollHeight; };
-    benchmarkES.addEventListener('done', function(e) {
-      benchmarkES.close();
-      benchmarkES = null;
-      term.textContent += (e.data ? '\nERROR: ' + e.data : '\nCompleted.') + '\n';
-      term.scrollTop = term.scrollHeight;
-      status.textContent = e.data ? 'Failed.' : 'Completed.';
-    });
+    const taskIds = benchmarkTaskIDs(d);
+    if (!taskIds.length) throw new Error('No benchmark task was queued.');
+    status.textContent = taskIds.length === 1 ? ('Task ' + taskIds[0] + ' queued.') : ('Queued ' + taskIds.length + ' tasks.');
+    const streamNext = function(idx, failures) {
+      if (idx >= taskIds.length) {
+        status.textContent = failures ? 'Completed with failures.' : 'Completed.';
+        return;
+      }
+      const taskId = taskIds[idx];
+      term.textContent += '\n[' + (idx + 1) + '/' + taskIds.length + '] Task ' + taskId + ' queued. Streaming log...\n';
+      benchmarkES = new EventSource('/api/tasks/' + taskId + '/stream');
+      benchmarkES.onmessage = function(e) { term.textContent += e.data + '\n'; term.scrollTop = term.scrollHeight; };
+      benchmarkES.addEventListener('done', function(e) {
+        benchmarkES.close();
+        benchmarkES = null;
+        if (e.data) failures += 1;
+        term.textContent += (e.data ? '\nERROR: ' + e.data : '\nCompleted.') + '\n';
+        term.scrollTop = term.scrollHeight;
+        streamNext(idx + 1, failures);
+      });
+      benchmarkES.onerror = function() {
+        if (benchmarkES) {
+          benchmarkES.close();
+          benchmarkES = null;
+        }
+        term.textContent += '\nERROR: stream disconnected.\n';
+        term.scrollTop = term.scrollHeight;
+        streamNext(idx + 1, failures + 1);
+      };
+    };
+    streamNext(0, 0);
   }).catch(function(err) {
     status.textContent = 'Error.';
     term.textContent += 'ERROR: ' + err.message + '\n';
@@ -1754,6 +1803,129 @@ benchmarkLoadGPUs();
 </script>`
 }
 
+func renderBenchmarkResultsCard(exportDir string) string {
+	columns, runs := loadBenchmarkHistory(exportDir)
+	return renderBenchmarkResultsCardFromRuns(
+		"Benchmark Results",
+		"Composite score by saved benchmark run and GPU.",
+		"No saved benchmark runs yet.",
+		columns,
+		runs,
+	)
+}
+
+func renderBenchmarkResultsCardFromRuns(title, description, emptyMessage string, columns []benchmarkHistoryColumn, 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>`
+	}
+	var b strings.Builder
+	b.WriteString(`<div class="card"><div class="card-head">` + html.EscapeString(title) + `</div><div class="card-body">`)
+	if strings.TrimSpace(description) != "" {
+		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(`</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 {
+				b.WriteString(`<td style="color:var(--muted)">-</td>`)
+				continue
+			}
+			b.WriteString(`<td>` + fmt.Sprintf("%.2f", cell.score) + `</td>`)
+		}
+		b.WriteString(`</tr>`)
+	}
+	b.WriteString(`</tbody></table></div></div></div>`)
+	return b.String()
+}
+
+func loadBenchmarkHistory(exportDir string) ([]benchmarkHistoryColumn, []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
+	}
+	sort.Strings(paths)
+	return loadBenchmarkHistoryFromPaths(paths)
+}
+
+func loadBenchmarkHistoryFromPaths(paths []string) ([]benchmarkHistoryColumn, []benchmarkHistoryRun) {
+	columnByKey := make(map[string]benchmarkHistoryColumn)
+	runs := make([]benchmarkHistoryRun, 0, len(paths))
+	for _, path := range paths {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			continue
+		}
+		var result platform.NvidiaBenchmarkResult
+		if err := json.Unmarshal(raw, &result); err != nil {
+			continue
+		}
+		run := benchmarkHistoryRun{
+			generatedAt: result.GeneratedAt,
+			displayTime: result.GeneratedAt.Local().Format("2006-01-02 15:04:05"),
+			cells:       make(map[string]benchmarkHistoryCell),
+		}
+		for _, gpu := range result.GPUs {
+			key := benchmarkHistoryColumnKey(gpu.Name, gpu.Index)
+			columnByKey[key] = benchmarkHistoryColumn{
+				key:   key,
+				label: benchmarkHistoryColumnLabel(gpu.Name, gpu.Index),
+				name:  strings.TrimSpace(gpu.Name),
+				index: gpu.Index,
+			}
+			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)
+	}
+	sort.Slice(columns, func(i, j int) bool {
+		leftName := strings.ToLower(strings.TrimSpace(columns[i].name))
+		rightName := strings.ToLower(strings.TrimSpace(columns[j].name))
+		if leftName != rightName {
+			return leftName < rightName
+		}
+		if columns[i].index != columns[j].index {
+			return columns[i].index < columns[j].index
+		}
+		return columns[i].key < columns[j].key
+	})
+	sort.Slice(runs, func(i, j int) bool {
+		return runs[i].generatedAt.After(runs[j].generatedAt)
+	})
+	return columns, runs
+}
+
+func benchmarkHistoryColumnKey(name string, index int) string {
+	return strings.TrimSpace(name) + "|" + strconv.Itoa(index)
+}
+
+func benchmarkHistoryColumnLabel(name string, index int) string {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return fmt.Sprintf("GPU %d", index)
+	}
+	return fmt.Sprintf("%s / GPU %d", name, index)
+}
+
 // ── Burn ──────────────────────────────────────────────────────────────────────
 
 func renderBurn() string {
@@ -1873,6 +2045,12 @@ func renderBurn() string {
 <script>
 let biES = null;
 
+function burnTaskIDs(payload) {
+  if (payload && Array.isArray(payload.task_ids) && payload.task_ids.length) return payload.task_ids;
+  if (payload && payload.task_id) return [payload.task_id];
+  return [];
+}
+
 function burnProfile() {
   const selected = document.querySelector('input[name="burn-profile"]:checked');
   return selected ? selected.value : 'smoke';
@@ -1974,6 +2152,9 @@ function streamTask(taskId, label) {
   });
 }
 function streamBurnTask(taskId, label, resetTerminal) {
+  return streamBurnTaskSet([taskId], label, resetTerminal);
+}
+function streamBurnTaskSet(taskIds, label, resetTerminal) {
   if (biES) { biES.close(); biES = null; }
   document.getElementById('bi-output').style.display = 'block';
   document.getElementById('bi-title').textContent = '— ' + label + ' [' + burnProfile() + ']';
@@ -1981,27 +2162,40 @@ function streamBurnTask(taskId, label, resetTerminal) {
   if (resetTerminal) {
     term.textContent = '';
   }
-  term.textContent += 'Task ' + taskId + ' queued. Streaming...\n';
-  return new Promise(function(resolve) {
-    biES = new EventSource('/api/tasks/' + taskId + '/stream');
-    biES.onmessage = function(e) { term.textContent += e.data + '\n'; term.scrollTop = term.scrollHeight; };
-    biES.addEventListener('done', function(e) {
-      biES.close();
-      biES = null;
-      term.textContent += (e.data ? '\nERROR: ' + e.data : '\nCompleted.') + '\n';
-      term.scrollTop = term.scrollHeight;
-      resolve({ok: !e.data, error: e.data || ''});
-    });
-    biES.onerror = function() {
-      if (biES) {
+  if (!Array.isArray(taskIds) || !taskIds.length) {
+    term.textContent += 'ERROR: no tasks queued.\n';
+    return Promise.resolve({ok:false, error:'no tasks queued'});
+  }
+  const streamNext = function(idx, failures) {
+    if (idx >= taskIds.length) {
+      return Promise.resolve({ok: failures === 0, error: failures ? (failures + ' task(s) failed') : ''});
+    }
+    const taskId = taskIds[idx];
+    term.textContent += '[' + (idx + 1) + '/' + taskIds.length + '] Task ' + taskId + ' queued. Streaming...\n';
+    return new Promise(function(resolve) {
+      biES = new EventSource('/api/tasks/' + taskId + '/stream');
+      biES.onmessage = function(e) { term.textContent += e.data + '\n'; term.scrollTop = term.scrollHeight; };
+      biES.addEventListener('done', function(e) {
         biES.close();
         biES = null;
-      }
-      term.textContent += '\nERROR: stream disconnected.\n';
-      term.scrollTop = term.scrollHeight;
-      resolve({ok: false, error: 'stream disconnected'});
-    };
-  });
+        term.textContent += (e.data ? '\nERROR: ' + e.data : '\nCompleted.') + '\n';
+        term.scrollTop = term.scrollHeight;
+        resolve(failures + (e.data ? 1 : 0));
+      });
+      biES.onerror = function() {
+        if (biES) {
+          biES.close();
+          biES = null;
+        }
+        term.textContent += '\nERROR: stream disconnected.\n';
+        term.scrollTop = term.scrollHeight;
+        resolve(failures + 1);
+      };
+    }).then(function(nextFailures) {
+      return streamNext(idx + 1, nextFailures);
+    });
+  };
+  return streamNext(0, 0);
 }
 
 function runBurnTaskSet(tasks, statusElId) {
@@ -2029,7 +2223,7 @@ function runBurnTaskSet(tasks, statusElId) {
     if (status) status.textContent = 'Running ' + (idx + 1) + '/' + enabled.length + '...';
     return enqueueBurnTask(t.target, t.label, t.extra, !!t.nvidia)
       .then(function(d) {
-        return streamBurnTask(d.task_id, t.label, false);
+        return streamBurnTaskSet(burnTaskIDs(d), t.label, false);
       })
       .then(function() {
         return runNext(idx + 1);

audit/internal/webui/server_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -636,6 +637,66 @@ 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")
+	if err := os.MkdirAll(runDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+			{
+				Index: 1,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1168.50,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{ExportDir: exportDir})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/benchmark", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	wantTime := result.GeneratedAt.Local().Format("2006-01-02 15:04:05")
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score by saved benchmark run and GPU.`,
+		`NVIDIA H100 PCIe / GPU 0`,
+		`NVIDIA H100 PCIe / GPU 1`,
+		`#1`,
+		wantTime,
+		`1176.25`,
+		`1168.50`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("benchmark page missing %q: %s", needle, body)
+		}
+	}
+}
+
 func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()

audit/internal/webui/task_page.go

@@ -97,10 +97,13 @@ func renderTaskDetailPage(opts HandlerOptions, task Task) string {
 		body.WriteString(`</div></div>`)
 	}
 
-	if task.Status == TaskRunning || task.Status == TaskPending {
+	if task.Status == TaskRunning {
 		body.WriteString(`<div class="card"><div class="card-head">Live Charts</div><div class="card-body">`)
 		body.WriteString(`<div id="task-live-charts" style="display:flex;flex-direction:column;gap:16px;color:var(--muted);font-size:13px">Loading charts...</div>`)
 		body.WriteString(`</div></div>`)
+	}
+
+	if task.Status == TaskRunning || task.Status == TaskPending {
 		body.WriteString(`<div class="card"><div class="card-head">Live Logs</div><div class="card-body">`)
 		body.WriteString(`<div id="task-live-log" class="terminal" style="max-height:none;white-space:pre-wrap">Connecting...</div>`)
 		body.WriteString(`</div></div>`)

audit/internal/webui/task_report.go

@@ -230,6 +230,9 @@ func renderTaskReportFragment(report taskReport, charts map[string]string, logTe
 	b.WriteString(`<div style="margin-top:14px;font-size:13px;color:var(--muted)">`)
 	b.WriteString(`Started: ` + formatTaskTime(report.StartedAt, report.CreatedAt) + ` | Finished: ` + formatTaskTime(report.DoneAt, time.Time{}) + ` | Duration: ` + formatTaskDuration(report.DurationSec))
 	b.WriteString(`</div></div></div>`)
+	if benchmarkCard := renderTaskBenchmarkResultsCard(report.Target, logText); benchmarkCard != "" {
+		b.WriteString(benchmarkCard)
+	}
 
 	if len(report.Charts) > 0 {
 		for _, chart := range report.Charts {
@@ -247,6 +250,57 @@ func renderTaskReportFragment(report taskReport, charts map[string]string, logTe
 	return b.String()
 }
 
+func renderTaskBenchmarkResultsCard(target, logText string) string {
+	if strings.TrimSpace(target) != "nvidia-benchmark" {
+		return ""
+	}
+	resultPath := taskBenchmarkResultPath(logText)
+	if strings.TrimSpace(resultPath) == "" {
+		return ""
+	}
+	columns, runs := loadBenchmarkHistoryFromPaths([]string{resultPath})
+	if len(runs) == 0 {
+		return ""
+	}
+	return renderBenchmarkResultsCardFromRuns(
+		"Benchmark Results",
+		"Composite score for this benchmark task.",
+		"No benchmark results were saved for this task.",
+		columns,
+		runs,
+	)
+}
+
+func taskBenchmarkResultPath(logText string) string {
+	archivePath := taskArchivePathFromLog(logText)
+	if archivePath == "" {
+		return ""
+	}
+	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
+	if runDir == archivePath {
+		return ""
+	}
+	return filepath.Join(runDir, "result.json")
+}
+
+func taskArchivePathFromLog(logText string) string {
+	lines := strings.Split(logText, "\n")
+	for i := len(lines) - 1; i >= 0; i-- {
+		line := strings.TrimSpace(lines[i])
+		if line == "" || !strings.HasPrefix(line, "Archive:") {
+			continue
+		}
+		path := strings.TrimSpace(strings.TrimPrefix(line, "Archive:"))
+		if strings.HasPrefix(path, "Archive written to ") {
+			path = strings.TrimSpace(strings.TrimPrefix(path, "Archive written to "))
+		}
+		if strings.HasSuffix(path, ".tar.gz") {
+			return path
+		}
+	}
+	return ""
+}
+
 func renderTaskStatusBadge(status string) string {
 	className := map[string]string{
 		TaskRunning:   "badge-ok",

audit/internal/webui/tasks.go

@@ -1149,7 +1149,32 @@ func taskArtifactsDir(root string, t *Task, status string) string {
 	if strings.TrimSpace(root) == "" || t == nil {
 		return ""
 	}
-	return filepath.Join(root, fmt.Sprintf("%s_%s_%s", t.ID, sanitizeTaskFolderPart(t.Name), taskFolderStatus(status)))
+	prefix := taskFolderNumberPrefix(t.ID)
+	return filepath.Join(root, fmt.Sprintf("%s_%s_%s", prefix, sanitizeTaskFolderPart(t.Name), taskFolderStatus(status)))
+}
+
+func taskFolderNumberPrefix(taskID string) string {
+	taskID = strings.TrimSpace(taskID)
+	if strings.HasPrefix(taskID, "TASK-") && len(taskID) >= len("TASK-000") {
+		num := strings.TrimSpace(strings.TrimPrefix(taskID, "TASK-"))
+		if len(num) == 3 {
+			allDigits := true
+			for _, r := range num {
+				if r < '0' || r > '9' {
+					allDigits = false
+					break
+				}
+			}
+			if allDigits {
+				return num
+			}
+		}
+	}
+	fallback := sanitizeTaskFolderPart(taskID)
+	if fallback == "" {
+		return "000"
+	}
+	return fallback
 }
 
 func ensureTaskReportPaths(t *Task) {

audit/internal/webui/tasks_test.go

@@ -163,6 +163,40 @@ func TestTaskQueueSnapshotSortsNewestFirst(t *testing.T) {
 	}
 }
 
+func TestNewJobIDUsesTASKPrefixAndZeroPadding(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	origCounter := jobCounter.Load()
+	jobCounter.Store(0)
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+		jobCounter.Store(origCounter)
+	})
+
+	if got := newJobID("ignored"); got != "TASK-000" {
+		t.Fatalf("id=%q want TASK-000", got)
+	}
+	if got := newJobID("ignored"); got != "TASK-001" {
+		t.Fatalf("id=%q want TASK-001", got)
+	}
+}
+
+func TestTaskArtifactsDirStartsWithTaskNumber(t *testing.T) {
+	root := t.TempDir()
+	task := &Task{
+		ID:   "TASK-007",
+		Name: "NVIDIA Benchmark",
+	}
+	got := filepath.Base(taskArtifactsDir(root, task, TaskDone))
+	if !strings.HasPrefix(got, "007_") {
+		t.Fatalf("artifacts dir=%q want prefix 007_", got)
+	}
+}
+
 func TestHandleAPITasksStreamReplaysPersistedLogWithoutLiveJob(t *testing.T) {
 	dir := t.TempDir()
 	logPath := filepath.Join(dir, "task.log")
@@ -325,6 +359,78 @@ func TestFinalizeTaskRunCreatesReportFolderAndArtifacts(t *testing.T) {
 	}
 }
 
+func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	benchmarkDir := filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	if err := os.MkdirAll(benchmarkDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(benchmarkDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	artifactsDir := filepath.Join(dir, "tasks", "task-bench_done")
+	if err := os.MkdirAll(artifactsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	task := &Task{
+		ID:           "task-bench",
+		Name:         "NVIDIA Benchmark",
+		Target:       "nvidia-benchmark",
+		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"
+	if err := os.WriteFile(task.LogPath, []byte(logText), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := writeTaskReportArtifacts(task); err != nil {
+		t.Fatalf("writeTaskReportArtifacts: %v", err)
+	}
+
+	body, err := os.ReadFile(task.ReportHTMLPath)
+	if err != nil {
+		t.Fatalf("ReadFile(report.html): %v", err)
+	}
+	html := string(body)
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score for this benchmark task.`,
+		`NVIDIA H100 PCIe / GPU 0`,
+		`1176.25`,
+	} {
+		if !strings.Contains(html, needle) {
+			t.Fatalf("report missing %q: %s", needle, html)
+		}
+	}
+}
+
 func TestTaskLifecycleMirrorsToSerialConsole(t *testing.T) {
 	var lines []string
 	prev := taskSerialWriteLine

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

@@ -62,6 +62,8 @@ done
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
@@ -78,7 +80,8 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
         fi
     fi
     echo "starting gpu ${id} size=${gpu_size_mb}MB"
-    "${WORKER}" --device "${id}" --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+    CUDA_VISIBLE_DEVICES="${id}" \
+        "${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
 done

iso/overlay/usr/local/bin/bee-john-gpu-stress

@@ -152,14 +152,19 @@ done
 
 [ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+export CUDA_VISIBLE_DEVICES="${FINAL}"
+
 JOHN_DEVICES=""
+local_id=1
 for id in $(echo "${FINAL}" | tr ',' ' '); do
-    opencl_id=$((id + 1))
+    opencl_id="${local_id}"
     if [ -z "${JOHN_DEVICES}" ]; then
         JOHN_DEVICES="${opencl_id}"
     else
         JOHN_DEVICES="${JOHN_DEVICES},${opencl_id}"
     fi
+    local_id=$((local_id + 1))
 done
 
 echo "loader=john"

iso/overlay/usr/local/bin/bee-nccl-gpu-stress

@@ -70,6 +70,8 @@ echo "gpu_count=${GPU_COUNT}"
 echo "range=${MIN_BYTES}..${MAX_BYTES}"
 echo "iters=${ITERS}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 deadline=$(( $(date +%s) + SECONDS ))
 round=0
 

iso/overlay/usr/local/bin/bee-nvidia-load

@@ -209,6 +209,18 @@ fi
 ldconfig 2>/dev/null || true
 log "ldconfig refreshed"
 
+# Keep persistence mode enabled across the session so dcgmi / stress tools do
+# not fail with deployment warnings on otherwise healthy GPUs.
+if command -v nvidia-smi >/dev/null 2>&1; then
+    if nvidia-smi -pm 1 >/dev/null 2>&1; then
+        log "enabled NVIDIA persistence mode"
+    else
+        log "WARN: failed to enable NVIDIA persistence mode"
+    fi
+else
+    log "WARN: nvidia-smi not found — cannot enable persistence mode"
+fi
+
 # Start DCGM host engine so dcgmi can discover GPUs.
 # nv-hostengine must run after the NVIDIA modules and device nodes are ready.
 # If it started too early (for example via systemd before bee-nvidia-load), it can