feat(ubenchx): add Huawei 950PR bandwidth / 新增 Huawei 950PR 带宽结果

docs/ubenchx-huawei-bandwidth.md

@@ -0,0 +1,13 @@
+# Huawei copy-bandwidth evidence
+
+The displayed series is **Huawei 950PR**, measured on Ascend950PR_9579 (28 Cube, 56 Vector, 128 GiB). The measured SKU remains in API metadata. This series does not change the global GPU specification catalog or inference-data normalization.
+
+`huawei-bandwidth.json` preserves all 42 full-precision plotted points, source input hashes, per-process medians, observed p10/p90 ranges, pool geometry and sample counts. Each size has three processes × six blocks × twenty observations. Aggregation is the pooled median amortized time, converted by `2 * message bytes / time`; bounds invert p90/p10 latency and are not confidence intervals. No observations are smoothed or supplied for missing sizes. The source is the validated October 7 rotating-pool 950PR experiment; the initial scratch-fill curves and overlap-rejected runs are excluded.
+
+The experiment uses FP16 `copy_` operations in complete rotating-pool NPU graph replays. Each direction has at least 512 MiB and two slots. Every block is checked against source bits before and after timing. The time is amortized over a complete rotation, not an isolated-copy latency. Unlike the GPU series, it is not `triton.testing.do_bench` on one FP32 buffer pair. The UI states this above the chart and labels Huawei tooltip times as amortized. No physical HBM-traffic counters prove that every access misses cache.
+
+The endpoint is about 1,447 GB/s at 16 GiB. MBU uses an assumed nominal reference of 1,600 GB/s; the exact board ceiling was not independently measured.
+
+UI/API coverage: the existing bandwidth/MBU/latency selector is presentation-only. `Huawei 950PR` is available through `test=mem-bw`, `gpu=<key>` and `gpu=all`; `test=all` returns empty sections for tests unavailable on an explicitly selected device. API rows retain optional uncertainty/pool fields, `metadata.protocol`, `dtype`, CANN/torch_npu versions and `missingSizesBytes`. `peakBandwidthSource` states the per-run reference qualification. The dashboard and API share `transformUbenchxRun`.
+
+Hover rulers use the currently selected metric and the same scaled coordinates as the displayed point, including after zoom.

packages/app/cypress/e2e/ubenchx.cy.ts

@@ -125,3 +125,93 @@ describe('ubenchX hub and test routes', () => {
     cy.get('meta[name="robots"]').should('have.attr', 'content').and('contain', 'noindex');
   });
 });
+
+const dismiss = {
+  onBeforeLoad(win: Cypress.AUTWindow) {
+    win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+  },
+};
+
+describe('Huawei copy-bandwidth evidence', () => {
+  it('renders Huawei 950PR and qualifies the graph timing in English and Chinese', () => {
+    cy.visit('/ubenchx/mem-bw', dismiss);
+    cy.contains('Huawei 950PR').should('be.visible');
+    cy.contains('910B3').should('not.exist');
+    cy.contains('Graph time is amortized per copy').should('be.visible');
+    cy.get('[data-testid="ubenchx-metric-select"]').click();
+    cy.contains('[role="option"]', 'Bandwidth (TB/s)').click();
+    cy.contains('h2', 'Bandwidth vs Message Size').should('be.visible');
+    cy.screenshot('huawei-bandwidth-en', { capture: 'fullPage' });
+    cy.visit('/zh/ubenchx/mem-bw', dismiss);
+    cy.contains('图执行时间按拷贝次数摊销').should('be.visible');
+    cy.contains('910B3').should('not.exist');
+    cy.screenshot('huawei-bandwidth-zh', { capture: 'fullPage' });
+  });
+  it('serves measured endpoints and empty unavailable tests through the public API', () => {
+    cy.request('/api/v1/views/ubenchx?test=all&gpu=Huawei%20950PR').then(({ body }) => {
+      expect(body.memBw.gpus[0].rows.at(-1).bytes).to.equal(17179869184);
+      expect(body.memBw.gpus[0].rows.at(-1).mbuPercent).to.be.closeTo(90.4545974, 0.0001);
+      expect(body.smL2Distance.gpus).to.deep.equal([]);
+      expect(body.tpcSkyline.gpus).to.deep.equal([]);
+    });
+  });
+});
+
+function assertBandwidthRuler(chart: string) {
+  cy.get(`${chart} circle.point`)
+    .filter((_, element) => {
+      const datum = (
+        element as unknown as SVGCircleElement & { __data__: { gpuKey: string; bytes: number } }
+      ).__data__;
+      return datum.gpuKey === 'H200 SXM' && datum.bytes === 33554432;
+    })
+    .should('have.length', 1)
+    .then(($point) => {
+      const x = Number($point.attr('cx')),
+        y = Number($point.attr('cy'));
+      cy.wrap($point).trigger('mouseenter', { force: true });
+      cy.get(`${chart} .horizontal-ruler`).should(($line) => {
+        expect(Number($line.attr('y1'))).to.be.closeTo(y, 0.01);
+        expect(Number($line.attr('y2'))).to.be.closeTo(y, 0.01);
+      });
+      cy.get(`${chart} .vertical-ruler`).should(($line) => {
+        expect(Number($line.attr('x1'))).to.be.closeTo(x, 0.01);
+        expect(Number($line.attr('x2'))).to.be.closeTo(x, 0.01);
+      });
+    });
+}
+
+it('keeps bandwidth hover guides on the rendered point in every metric and after zoom', () => {
+  cy.visit('/ubenchx/mem-bw', dismiss);
+  assertBandwidthRuler('#ubenchx-mbu');
+  for (const [label, chart] of [
+    ['Bandwidth (TB/s)', '#ubenchx-bandwidth'],
+    ['Latency (ms)', '#ubenchx-latency'],
+  ]) {
+    cy.get('[data-testid="ubenchx-metric-select"]').click();
+    cy.contains('[role="option"]', label).click();
+    assertBandwidthRuler(chart);
+  }
+  cy.get('#ubenchx-latency circle.point')
+    .first()
+    .then(($point) => {
+      const previousX = $point.attr('cx');
+      cy.get('#ubenchx-latency svg').then(($svg) => {
+        const rect = $svg[0].getBoundingClientRect();
+        cy.wrap($svg).trigger('wheel', {
+          eventConstructor: 'WheelEvent',
+          deltaY: -200,
+          shiftKey: true,
+          clientX: rect.left + 150,
+          clientY: rect.top + 100,
+          force: true,
+        });
+      });
+      cy.get('#ubenchx-latency circle.point')
+        .first()
+        .should(($current) => {
+          expect($current.attr('cx')).not.to.equal(previousX);
+        });
+    });
+  assertBandwidthRuler('#ubenchx-latency');
+});

packages/app/src/app/api/v1/views/ubenchx/route.test.ts

@@ -158,3 +158,47 @@ describe('ubenchx view', () => {
     expect(h200Last.mbuPercent).toBeCloseTo((4305.39 / 4800) * 100, 4);
   });
 });
+
+it('exposes Huawei rotating-copy evidence without inventing missing tests or sizes', async () => {
+  const response = await GET(request({ test: 'all', gpu: 'huawei 950pr' }));
+  expect(response.status).toBe(200);
+  const body = await response.json();
+  const run = body.memBw.gpus[0];
+  expect(run.gpu).toBe('Huawei 950PR');
+  expect(run.peakBandwidthGbps).toBe(1600);
+  expect(run.peakBandwidthSource).toContain('not independently verified');
+  expect(run.metadata).toMatchObject({
+    protocol: 'rotating-pool-npu-graph',
+    dtype: 'float16',
+    missingSizesBytes: [],
+  });
+  expect(run.rows.some((row: { bytes: number }) => row.bytes === 17179869184)).toBe(true);
+  const endpoint = run.rows.find((row: { bytes: number }) => row.bytes === 17179869184);
+  expect(endpoint.bandwidthGbps).toBeCloseTo(1447.2735584561165, 6);
+  expect(endpoint.mbuPercent).toBeCloseTo(90.45459740350728, 6);
+  expect(endpoint.observations).toBe(360);
+  expect(body.smL2Distance.gpus).toEqual([]);
+  expect(body.tpcSkyline.gpus).toEqual([]);
+});
+
+it('preserves full-precision Huawei bandwidth arithmetic and API/UI normalization', async () => {
+  for (const gpu of ['Huawei 950PR']) {
+    const response = await GET(request({ test: 'mem-bw', gpu }));
+    const body = await response.json();
+    const run = body.gpus[0];
+    expect(run).toEqual(transformUbenchxRun(gpu, UBENCHX_RUNS[gpu]));
+    for (const row of run.rows) {
+      expect(row.bandwidthGbps).toBeCloseTo((2 * row.bytes) / row.timeMs / 1e6, 6);
+      expect(row.bandwidthP10Gbps).toBeLessThanOrEqual(row.bandwidthGbps);
+      expect(row.bandwidthP90Gbps).toBeGreaterThanOrEqual(row.bandwidthGbps);
+    }
+  }
+});
+
+it('does not expose the removed 910B3 bandwidth series', async () => {
+  const all = await GET(request({ test: 'mem-bw', gpu: 'all' }));
+  const body = await all.json();
+  expect(body.gpus.some((run: { gpu: string }) => run.gpu === 'Ascend 910B3')).toBe(false);
+  const removed = await GET(request({ test: 'mem-bw', gpu: 'Ascend 910B3' }));
+  expect(removed.status).toBe(400);
+});

packages/app/src/app/api/v1/views/ubenchx/route.ts

@@ -64,13 +64,11 @@ export function GET(request: NextRequest) {
       }
 
       // test === 'all': return every test
-      const smL2GpuKeys = Object.keys(SM_L2_RUNS);
-      const smL2Gpu = parseEnumParam(s.get('gpu'), 'gpu', ['all', ...smL2GpuKeys], 'all');
       const smL2Results =
-        smL2Gpu === 'all'
+        rawGpu === null || gpu === 'all'
           ? Object.entries(SM_L2_RUNS).map(([key, run]) => transformSmL2Run(key, run))
-          : SM_L2_RUNS[smL2Gpu]
-            ? [transformSmL2Run(smL2Gpu, SM_L2_RUNS[smL2Gpu])]
+          : SM_L2_RUNS[gpu]
+            ? [transformSmL2Run(gpu, SM_L2_RUNS[gpu])]
             : [];
 
       return Promise.resolve(
@@ -85,11 +83,12 @@ export function GET(request: NextRequest) {
           // Like smL2Distance, default to every GPU rather than mem-bw's first one, and
           // match the GPU key case-insensitively as parseEnumParam does for the others.
           tpcSkyline: {
-            gpus: tpcSkylineResults(
-              Object.keys(TPC_SKYLINE_RUNS).find(
-                (key) => key.toLowerCase() === rawGpu?.toLowerCase(),
-              ) ?? 'all',
-            ),
+            gpus:
+              rawGpu === null || gpu === 'all'
+                ? tpcSkylineResults('all')
+                : TPC_SKYLINE_RUNS[gpu]
+                  ? tpcSkylineResults(gpu)
+                  : [],
           },
         }),
       );

packages/app/src/components/ubenchx/UbenchxContent.tsx

@@ -16,7 +16,12 @@ import { useLocale } from '@/lib/i18n/use-locale';
 import type { Locale } from '@/lib/i18n/i18n';
 
 import { UBENCHX_GPU_KEYS, UBENCHX_RUNS } from './ubenchx-data';
-import { transformUbenchxRun, formatBytes, type UbenchxDerivedRow } from './ubenchx-transform';
+import {
+  transformUbenchxRun,
+  formatBytes,
+  sizeAxisTicks,
+  type UbenchxDerivedRow,
+} from './ubenchx-transform';
 
 const GPU_COLORS: Record<string, string> = {
   'B300 SXM': 'oklch(0.62 0.15 128)',
@@ -27,13 +32,16 @@ const GPU_COLORS: Record<string, string> = {
   MI355X: 'oklch(0.62 0.20 18)',
   MI325X: 'oklch(0.55 0.20 28)',
   MI300X: 'oklch(0.48 0.20 38)',
+  'Huawei 950PR': '#C7000B',
 };
 
 const STRINGS = {
   en: {
     pageTitle: 'HBM Bandwidth',
     pageSubtitle:
-      'Microbenchmark measuring device-memory copy bandwidth across message sizes on NVIDIA and AMD GPUs.',
+      'Microbenchmark measuring device-memory copy bandwidth across message sizes on NVIDIA, AMD and Huawei accelerators.',
+    protocolNote:
+      'Huawei uses amortized rotating-buffer graph timing; NVIDIA/AMD use Triton copy timing. Small-message latency is not directly comparable.',
     latencyTitle: 'Latency vs Message Size',
     latencyY: 'Latency (ms)',
     mbuTitle: 'Memory Bandwidth Utilization (MBU) vs Message Size',
@@ -42,10 +50,16 @@ const STRINGS = {
     bwY: 'Bandwidth (TB/s)',
     xAxis: 'Message Size',
     yMetric: 'Y-axis',
-    peakNote: "MBU is relative to each GPU's own peak HBM bandwidth from GPU_SPECS.",
+    peakNote:
+      'MBU uses each device’s bandwidth reference. Huawei’s 1.6 TB/s references are qualified below.',
     methodology: 'Methodology',
     methodologyText:
-      'Times b.copy_(a) on float32 tensors using triton.testing.do_bench; bandwidth = 2 * bytes / time (read + write). Power-of-two sizes from 8 B to 16 GiB.',
+      'NVIDIA/AMD: times b.copy_(a) on float32 tensors using triton.testing.do_bench; bandwidth = 2 * bytes / time (read + write). Power-of-two sizes from 8 B to 16 GiB.',
+    huaweiMethodology:
+      'Huawei: FP16 rotating source/destination pools of at least 512 MiB each, timed as full rotating-buffer NPU graph replays. Three processes, 360 observations per size; full copied-bit checks before/after each block. Graph time is amortized per copy, not isolated-copy latency. Protocols differ from the GPU curves. No physical HBM-traffic counters were collected.',
+    huaweiReference:
+      'Huawei MBU uses a 1,600 GB/s reference; the exact board bandwidth ceiling was not independently measured.',
+    amortized: 'Amortized graph copy time',
     source: 'Source',
     instructions:
       'Shift+Scroll to zoom · Drag to pan · Double-click to reset · Click a point to pin tooltip',
@@ -54,12 +68,14 @@ const STRINGS = {
     tooltipLatency: 'Latency',
     tooltipBandwidth: 'Bandwidth',
     tooltipMbu: 'MBU',
-    tooltipPeak: 'Peak BW',
+    tooltipPeak: 'Bandwidth reference',
     aria: 'ubenchX device-memory copy bandwidth charts',
   },
   zh: {
     pageTitle: 'HBM 带宽',
-    pageSubtitle: '在 NVIDIA 和 AMD GPU 上测量不同消息大小的显存拷贝带宽微基准测试。',
+    pageSubtitle: '测量 NVIDIA、AMD 和 Huawei 加速器在不同消息大小下的显存拷贝带宽。',
+    protocolNote:
+      'Huawei 使用轮转缓冲区的图执行摊销计时；NVIDIA/AMD 使用 Triton 拷贝计时。小消息延迟不能直接比较。',
     latencyTitle: '延迟 vs 消息大小',
     latencyY: '延迟（ms）',
     mbuTitle: '显存带宽利用率（MBU）vs 消息大小',
@@ -68,18 +84,22 @@ const STRINGS = {
     bwY: '带宽（TB/s）',
     xAxis: '消息大小',
     yMetric: 'Y 轴',
-    peakNote: 'MBU 基于各 GPU 自身在 GPU_SPECS 中的峰值 HBM 带宽计算。',
+    peakNote: 'MBU 按各设备的带宽参考值计算。Huawei 的 1.6 TB/s 参考值及其限制见下文。',
     methodology: '测试方法',
     methodologyText:
-      '使用 triton.testing.do_bench 计时 b.copy_(a)（float32 张量）；带宽 = 2 * bytes / time（读 + 写）。消息大小为 2 的幂次，从 8 B 到 16 GiB。',
+      'NVIDIA/AMD：使用 triton.testing.do_bench 计时 b.copy_(a)（float32 张量）；带宽 = 2 * bytes / time（读 + 写）。消息大小为 2 的幂次，从 8 B 到 16 GiB。',
+    huaweiMethodology:
+      'Huawei：源和目标分别使用至少 512 MiB 的 FP16 轮转缓冲池，每次计时覆盖完整缓冲池轮转的一次 NPU 图重放。每个大小在三个进程中共测量 360 次，每个测量块前后逐位校验全部拷贝结果。图执行时间按拷贝次数摊销，不代表单次独立拷贝延迟；与 GPU 曲线的测试协议不同，且未采集物理 HBM 流量计数器。',
+    huaweiReference: 'Huawei 的 MBU 使用 1,600 GB/s 参考值，该板卡的确切带宽上限未经独立实测。',
+    amortized: '摊销后的图拷贝时间',
     source: '数据来源',
     instructions: 'Shift+滚轮缩放 · 拖动平移 · 双击重置 · 点击数据点固定提示框',
     dismiss: '点击其他区域关闭',
     tooltipSize: '大小',
     tooltipLatency: '延迟',
     tooltipBandwidth: '带宽',
     tooltipMbu: 'MBU',
-    tooltipPeak: '峰值带宽',
+    tooltipPeak: '带宽参考值',
     aria: 'ubenchX 显存拷贝带宽图表',
   },
 } as const;
@@ -143,7 +163,7 @@ function formatTooltip(point: ChartPoint, isPinned: boolean, t: (typeof STRINGS)
         <span style="display: inline-block; width: 8px; height: 8px; border-radius: 9999px; background: ${color};"></span>${point.gpuKey}
       </div>
       ${tooltipLine(t.tooltipSize, point.sizeLabel)}
-      ${tooltipLine(t.tooltipLatency, `${point.timeMs.toFixed(4)} ms`)}
+      ${tooltipLine(UBENCHX_RUNS[point.gpuKey].metadata.protocol ? t.amortized : t.tooltipLatency, `${point.timeMs.toFixed(4)} ms`)}
       ${tooltipLine(t.tooltipBandwidth, `${(point.bandwidthGbps / 1000).toFixed(3)} TB/s`)}
       ${tooltipLine(t.tooltipMbu, `${point.mbuPercent.toFixed(1)}%`)}
       ${tooltipLine(t.tooltipPeak, `${point.peakBandwidthGbps / 1000} TB/s`)}
@@ -215,10 +235,7 @@ function UbenchxChart({
     return ticks;
   }, [yScaleType, yDomain]);
 
-  const xTickValues = useMemo(() => {
-    const all = [...new Set(allPoints.map((p) => p.x))].sort((a, b) => a - b);
-    return all.filter((_, i) => i % 3 === 0 || i === all.length - 1);
-  }, [allPoints]);
+  const xTickValues = useMemo(() => sizeAxisTicks(...xDomain), [xDomain]);
 
   return (
     <D3Chart<ChartPoint>
@@ -284,6 +301,8 @@ function UbenchxChart({
       }}
       tooltip={{
         rulerType: 'crosshair',
+        getRulerX: (point, scale) => (scale as d3.ScaleLogarithmic<number, number>)(point.x),
+        getRulerY: (point, scale) => scale(point.y),
         attachToLayer: 1,
         content: (point, isPinned) => formatTooltip(point, isPinned, STRINGS[locale]),
       }}
@@ -324,6 +343,7 @@ export function UbenchxContent() {
       <div>
         <h2 className="text-xl font-semibold tracking-tight">{t.pageTitle}</h2>
         <p className="text-sm text-muted-foreground mt-1">{t.pageSubtitle}</p>
+        <p className="text-sm text-muted-foreground mt-2">{t.protocolNote}</p>
       </div>
 
       <div className="flex flex-col gap-1.5 sm:flex-row sm:items-center sm:gap-3">
@@ -376,11 +396,16 @@ export function UbenchxContent() {
       <section className="text-xs text-muted-foreground space-y-1">
         <h3 className="text-sm font-semibold">{t.methodology}</h3>
         <p>{t.methodologyText}</p>
+        <p>{t.huaweiMethodology}</p>
+        <p>{t.huaweiReference}</p>
         {viewResults.map((r) => (
           <p key={r.gpu}>
             <strong>{r.gpu}:</strong> {r.metadata.gpu} | Driver: {r.metadata.driver} | PyTorch:{' '}
-            {r.metadata.torch} | Triton: {r.metadata.triton} | Container: {r.metadata.container} |
-            Peak: {r.peakBandwidthGbps / 1000} TB/s (GPU_SPECS)
+            {r.metadata.torch} |{' '}
+            {r.metadata.cann
+              ? `CANN: ${r.metadata.cann} | torch_npu: ${r.metadata.torchNpu}`
+              : `Triton: ${r.metadata.triton} | Container: ${r.metadata.container}`}{' '}
+            |{r.peakBandwidthGbps / 1000} TB/s
           </p>
         ))}
         {sourceUrls.map((url) => (