SemiAnalysisAI · edwingao28 · Aug 27, 2026 · Aug 27, 2026 · Aug 27, 2026 · Aug 27, 2026
@@ -8,6 +8,7 @@ const availableQuickFilters: QuickFilters = {
   frameworks: ['vllm', 'sglang'],
   deployment: ['single-node', 'multi-node', 'disagg'],
   spec: ['mtp', 'stp'],
+  power: ['certified', 'legacy'],
 };
 
 describe('QuickFiltersDialog', () => {
@@ -47,6 +48,7 @@ describe('QuickFiltersDialog', () => {
           frameworks: ['sglang'],
           deployment: ['disagg'],
           spec: ['mtp'],
+          power: ['certified'],
         },
         availableQuickFilters,
       },
@@ -57,5 +59,6 @@ describe('QuickFiltersDialog', () => {
     cy.get('@setQuickFilterFrameworks').should('have.been.calledWith', []);
     cy.get('@setQuickFilterDeployment').should('have.been.calledWith', []);
     cy.get('@setQuickFilterSpec').should('have.been.calledWith', []);
+    cy.get('@setQuickFilterPower').should('have.been.calledWith', []);
   });
 });
@@ -0,0 +1,174 @@
+// Certified-vs-legacy measured power (PLAN-02 / gap G2): on the Measured
+// Energy y-axes, points without a producer validation verdict carry a dotted
+// ring and a footer legend key, and Quick Filters gains a "Measured Power"
+// category (Certified/Legacy pills, `i_power` share-link param). Fixture rows
+// are intercepted so one config is certified (power_valid=1) and one is
+// legacy (no verdict), keeping every assertion deterministic regardless of
+// what the production dataset contains.
+
+const POWER_MODEL = 'dsv4';
+const POWER_DATE = '2026-08-20';
+
+const powerPercentileLadder = (prefix: string, base: number): Record<string, number> => ({
+  [`median_${prefix}`]: base,
+  [`p75_${prefix}`]: base * 1.2,
+  [`p90_${prefix}`]: base * 1.5,
+  [`p95_${prefix}`]: base * 1.7,
+  [`p99_${prefix}`]: base * 2.2,
+  [`std_${prefix}`]: base * 0.3,
+});
+
+const powerMetrics = (conc: number): Record<string, number> => {
+  const scale = conc / 16;
+  const itl = 0.011 * scale;
+  return {
+    ...powerPercentileLadder('ttft', 0.4 * scale),
+    ...powerPercentileLadder('tpot', 0.012 * scale),
+    ...powerPercentileLadder('itl', itl),
+    ...powerPercentileLadder('e2el', 8 * scale),
+    median_intvty: 1 / itl,
+    p75_intvty: 1 / (itl * 1.2),
+    p90_intvty: 1 / (itl * 1.5),
+    p99_intvty: 1 / (itl * 2.2),
+    std_intvty: (1 / itl) * 0.1,
+    tput_per_gpu: 950 / Math.sqrt(scale),
+    output_tput_per_gpu: 210,
+    input_tput_per_gpu: 740,
+  };
+};
+
+const powerConfigs = [
+  // Legacy telemetry: measured watts without a producer verdict.
+  { hardware: 'b200', framework: 'vllm', power: { avg_power_w: 560 } },
+  // Certified telemetry: explicit power_valid=1 verdict.
+  { hardware: 'mi300x', framework: 'sglang', power: { power_valid: 1, avg_power_w: 685.5 } },
+];
+
+const powerAvailability = powerConfigs.map((config) => ({
+  model: POWER_MODEL,
+  isl: 8192,
+  osl: 1024,
+  precision: 'fp4',
+  hardware: config.hardware,
+  framework: config.framework,
+  spec_method: 'none',
+  disagg: false,
+  benchmark_type: 'single_turn',
+  date: POWER_DATE,
+}));
+
+let powerBenchmarkId = 990000;
+const powerBenchmarks = powerConfigs.flatMap((config) =>
+  [16, 64, 128].map((conc) => ({
+    id: powerBenchmarkId++,
+    hardware: config.hardware,
+    framework: config.framework,
+    model: POWER_MODEL,
+    precision: 'fp4',
+    spec_method: 'none',
+    disagg: false,
+    is_multinode: false,
+    prefill_tp: 8,
+    decode_tp: 8,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    isl: 8192,
+    osl: 1024,
+    conc,
+    offload_mode: 'off',
+    benchmark_type: 'single_turn',
+    image: `${config.framework}/server:test`,
+    metrics: { ...powerMetrics(conc), ...config.power },
+    workers: null,
+    date: POWER_DATE,
+    run_url: null,
+  })),
+);
+
+function visitCertifiedPowerChart(extraParams = '') {
+  cy.intercept('GET', '/api/v1/availability', { body: powerAvailability }).as('availability');
+  cy.intercept('GET', '/api/v1/benchmarks*', { body: powerBenchmarks }).as('benchmarks');
+  cy.visit(`/inference?g_model=DeepSeek-V4-Pro&i_seq=8k/1k&i_prec=fp4${extraParams}`, {
+    onBeforeLoad(win) {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    },
+  });
+  cy.wait(['@availability', '@benchmarks']);
+  cy.get('[data-testid="inference-chart-display"]', { timeout: 30_000 }).should('exist');
+  cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
+}
+
+describe('Certified vs legacy measured power', () => {
+  it('rings legacy points on a measured axis and filters them via Quick Filters', () => {
+    visitCertifiedPowerChart();
+
+    // No decorations off the Measured Energy axes.
+    cy.get('.legacy-power-ring').should('not.exist');
+    cy.get('[data-testid="legacy-power-key"]').should('not.exist');
+
+    // Pick "Measured Average Power per Chip" from the y-axis dropdown. The
+    // select list is a scroll container and Measured Energy sits below the
+    // fold, so scroll before clicking.
+    cy.get('[data-testid="yaxis-metric-selector"]').click();
+    cy.contains('[role="option"]', 'Measured Average Power per Chip')
+      .scrollIntoView()
+      .should('be.visible')
+      .click();
+    cy.get('[data-slot="select-content"]').should('not.exist');
+
+    // The no-verdict config is ringed; the certified one is not. The footer
+    // legend key appears with the ringed points.
+    cy.get('.dot-group[data-hw-key^="b200"] .legacy-power-ring').should('exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"] .legacy-power-ring').should('not.exist');
+    cy.get('[data-testid="legacy-power-key"]').should('be.visible');
+    cy.screenshot('legacy-power-rings', { capture: 'viewport' });
+
+    // Quick Filters gains the Measured Power category with both pills enabled.
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
+    cy.get('[data-testid="quick-filter-power-certified"]').should('be.enabled');
+    cy.get('[data-testid="quick-filter-power-legacy"]').should('be.enabled');
+
+    // Certified-only: legacy points (and with them every ring and the legend
+    // key) leave the chart while the certified series stays.
+    cy.get('[data-testid="quick-filter-power-certified"]').click();
+    cy.get('[data-testid="quick-filters-selected-count"]').should('contain.text', '1 selected');
+    cy.get('.dot-group[data-hw-key^="b200"]').should('not.exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+    cy.get('.legacy-power-ring').should('not.exist');
+    cy.get('[data-testid="legacy-power-key"]').should('not.exist');
+    cy.get('[data-testid="inference-chart-display"] svg').should('exist');
+    cy.screenshot('certified-only-filter', { capture: 'viewport' });
+
+    // Clear filters restores the legacy series, rings, and legend key.
+    cy.contains('button', 'Clear filters').click();
+    cy.get('[data-testid="quick-filters-selected-count"]').should('not.exist');
+    cy.get('[data-testid="quick-filter-power-certified"]').should(
+      'have.attr',
+      'aria-pressed',
+      'false',
+    );
+    cy.contains('button', 'Done').click();
+    cy.get('.dot-group[data-hw-key^="b200"] .legacy-power-ring').should('exist');
+    cy.get('[data-testid="legacy-power-key"]').should('be.visible');
+  });
+
+  it('restores a shared i_power=certified link with the pill pre-selected', () => {
+    // Note: filter writes live in the in-memory share-link store (the address
+    // bar is deliberately stripped after load — see url-state.ts), so the
+    // durable observable behavior is the restore direction tested here.
+    visitCertifiedPowerChart('&i_metric=y_measuredAvgPower&i_power=certified');
+
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+    cy.get('.dot-group[data-hw-key^="b200"]').should('not.exist');
+    cy.get('[data-testid="legacy-power-key"]').should('not.exist');
+
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filter-power-certified"]').should(
+      'have.attr',
+      'aria-pressed',
+      'true',
+    );
+    cy.get('[data-testid="quick-filters-selected-count"]').should('contain.text', '1 selected');
+  });
+});
@@ -230,12 +230,13 @@ export function createMockInferenceContextValues(
     setSelectedXAxisMode: namedStub('setSelectedXAxisMode'),
     scaleType: 'auto',
     setScaleType: namedStub('setScaleType'),
-    quickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
-    availableQuickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
+    quickFilters: { vendors: [], frameworks: [], deployment: [], spec: [], power: [] },
+    availableQuickFilters: { vendors: [], frameworks: [], deployment: [], spec: [], power: [] },
     setQuickFilterVendors: namedStub('setQuickFilterVendors'),
     setQuickFilterFrameworks: namedStub('setQuickFilterFrameworks'),
     setQuickFilterDeployment: namedStub('setQuickFilterDeployment'),
     setQuickFilterSpec: namedStub('setQuickFilterSpec'),
+    setQuickFilterPower: namedStub('setQuickFilterPower'),
     isLegendExpanded: true,
     setIsLegendExpanded: namedStub('setIsLegendExpanded'),
     hideNonOptimal: false,

@@ -95,7 +95,9 @@ import { bestSeriesPerSku } from './utils/best-series-per-sku';
 import {
   EMPTY_QUICK_FILTERS,
   parseDeploymentModes,
+  parsePowerTiers,
   type DeploymentMode,
+  type PowerTier,
   type QuickFilters,
   type SpecMode,
 } from './utils/quickFilters';
@@ -425,7 +427,7 @@ export function InferenceProvider({
     () => (getUrlParam('i_scale') as 'auto' | 'linear' | 'log') || 'auto',
   );
 
-  // ── Quick filters (vendor / framework / deployment / mtp-stp) ───────────────
+  // ── Quick filters (vendor / framework / deployment / mtp-stp / power tier) ──
   // Coarse pre-filters applied to the point set. Empty = no constraint.
   //
   // Initialized empty rather than from the URL so the first client render matches
@@ -438,8 +440,9 @@ export function InferenceProvider({
   const [quickFilterFrameworks, setQuickFilterFrameworks] = useState<string[]>([]);
   const [quickFilterDeployment, setQuickFilterDeployment] = useState<DeploymentMode[]>([]);
   const [quickFilterSpec, setQuickFilterSpec] = useState<SpecMode[]>([]);
+  const [quickFilterPower, setQuickFilterPower] = useState<PowerTier[]>([]);
   useEffect(() => {
-    const parse = (key: 'i_vendor' | 'i_fw' | 'i_disagg' | 'i_spec') => {
+    const parse = (key: 'i_vendor' | 'i_fw' | 'i_disagg' | 'i_spec' | 'i_power') => {
       const v = getUrlParam(key);
       return v ? v.split(',').filter(Boolean) : [];
     };
@@ -449,19 +452,28 @@ export function InferenceProvider({
     // point, so expand it to both aggregate deployment modes.
     const deployment = parseDeploymentModes(parse('i_disagg'));
     const spec = parse('i_spec') as SpecMode[];
+    const power = parsePowerTiers(parse('i_power'));
     if (vendors.length > 0) setQuickFilterVendors(vendors);
     if (frameworks.length > 0) setQuickFilterFrameworks(frameworks);
     if (deployment.length > 0) setQuickFilterDeployment(deployment);
     if (spec.length > 0) setQuickFilterSpec(spec);
+    if (power.length > 0) setQuickFilterPower(power);
   }, [getUrlParam]);
   const quickFilters = useMemo<QuickFilters>(
     () => ({
       vendors: quickFilterVendors,
       frameworks: quickFilterFrameworks,
       deployment: quickFilterDeployment,
       spec: quickFilterSpec,
+      power: quickFilterPower,
     }),
-    [quickFilterVendors, quickFilterFrameworks, quickFilterDeployment, quickFilterSpec],
+    [
+      quickFilterVendors,
+      quickFilterFrameworks,
+      quickFilterDeployment,
+      quickFilterSpec,
+      quickFilterPower,
+    ],
   );
   // Historical Trends hides Quick Filters, so never apply invisible selections there.
   // Agentic charts expose vendor, framework, and deployment filters, but speculative
@@ -1473,6 +1485,7 @@ export function InferenceProvider({
       i_fw: quickFilterFrameworks.join(','),
       i_disagg: quickFilterDeployment.join(','),
       i_spec: quickFilterSpec.join(','),
+      i_power: quickFilterPower.join(','),
     },
     [
       selectedYAxisMetric,
@@ -1499,6 +1512,7 @@ export function InferenceProvider({
       quickFilterFrameworks,
       quickFilterDeployment,
       quickFilterSpec,
+      quickFilterPower,
     ],
   );
 
@@ -1763,6 +1777,7 @@ export function InferenceProvider({
     setQuickFilterFrameworks,
     setQuickFilterDeployment,
     setQuickFilterSpec,
+    setQuickFilterPower,
     setIsLegendExpanded,
     setHideNonOptimal,
     setShowPointLabels,

@@ -197,6 +197,18 @@ function provisionedJoules(tokenType: TokenType): MetricExplanation {
   };
 }
 
+/**
+ * Certification-tier note appended to every Measured Energy explanation:
+ * pairs the dotted legacy-power point ring and the Quick Filters → Measured
+ * Power pills with the footer text that explains them.
+ */
+const MEASURED_TIER_NOTE_EN =
+  ' Points without a producer validation verdict are legacy measurements — marked with a ' +
+  'dotted ring on the chart and filterable via Quick Filters → Measured Power.';
+const MEASURED_TIER_NOTE_ZH =
+  '没有生产端验证结论的数据点属于旧版测量——在图表上以虚线圆环标记，' +
+  '并可通过快捷筛选中的“实测功耗”进行筛选。';
+
 type MeasuredPhase = 'run' | 'prefill' | 'decode';
 
 const MEASURED_PHASE_EN: Record<MeasuredPhase, string> = {
@@ -216,11 +228,13 @@ function measuredPower(phase: MeasuredPhase): MetricExplanation {
     description: {
       en:
         `Average per-chip accelerator power actually drawn during ${MEASURED_PHASE_EN[phase]}, ` +
-        'read from runner telemetry. Unlike the all-in provisioned metrics, this reflects real ' +
-        'measured draw, not the provisioned budget.',
+        `read from runner telemetry. Unlike the all-in provisioned metrics, this reflects real ` +
+        `measured draw, not the provisioned budget.${MEASURED_TIER_NOTE_EN}`,
       zh:
         `${MEASURED_PHASE_ZH[phase]}每块加速器芯片的实际平均功耗，来自运行器遥测数据。` +
-        '与全电源配置类指标不同，它反映的是真实实测功耗，而不是按配置计算的预算值。',
+        `与全电源配置类指标不同，它反映的是真实实测功耗，而不是按配置计算的预算值。${
+          MEASURED_TIER_NOTE_ZH
+        }`,
     },
     formula: {
       en: `W = mean of sampled per-chip accelerator power draw over ${MEASURED_PHASE_EN[phase]}`,
@@ -236,10 +250,10 @@ function measuredJoulesPerToken(tokenType: TokenType): MetricExplanation {
         `Measured accelerator energy consumed per ${
           tokenType === 'total' ? 'token (including prompt tokens)' : `${tokenType} token`
         }, from runner power telemetry integrated over the run. Lower means the system converts ` +
-        'electricity into tokens more efficiently.',
+        `electricity into tokens more efficiently.${MEASURED_TIER_NOTE_EN}`,
       zh:
         `每个${TOKEN_TYPE_ZH[tokenType]}消耗的加速器实测能耗，由运行器功耗遥测在整个运行期间积分得到。` +
-        '数值越低，说明系统把电能转化为 token 的效率越高。',
+        `数值越低，说明系统把电能转化为 token 的效率越高。${MEASURED_TIER_NOTE_ZH}`,
     },
     formula: {
       en: `J/tok = measured accelerator energy over the run ÷ ${TOKEN_TYPE_EN[tokenType]} processed`,
@@ -341,12 +355,12 @@ export const METRIC_EXPLANATIONS: Record<MetricKey, MetricExplanation> = {
   measuredJPerSuccessfulQuery: {
     description: {
       en:
-        'Measured accelerator energy consumed per successfully completed request, from runner ' +
-        'power telemetry. It charges the energy of the whole run only to requests that finished ' +
-        'successfully.',
+        `Measured accelerator energy consumed per successfully completed request, from runner ` +
+        `power telemetry. It charges the energy of the whole run only to requests that finished ` +
+        `successfully.${MEASURED_TIER_NOTE_EN}`,
       zh:
-        '每个成功完成的请求消耗的加速器实测能耗，来自运行器功耗遥测。' +
-        '整个运行的能耗只计入成功完成的请求。',
+        `每个成功完成的请求消耗的加速器实测能耗，来自运行器功耗遥测。` +
+        `整个运行的能耗只计入成功完成的请求。${MEASURED_TIER_NOTE_ZH}`,
     },
     formula: {
       en: 'J/query = measured accelerator energy over the run ÷ successfully completed requests',
@@ -356,9 +370,9 @@ export const METRIC_EXPLANATIONS: Record<MetricKey, MetricExplanation> = {
   measuredWhPerSuccessfulQuery: {
     description: {
       en:
-        'The same measured energy per successful request expressed in watt-hours, a more ' +
-        'familiar household unit (1 Wh = 3,600 J).',
-      zh: '与每次成功请求实测能耗相同的量，换算成更直观的瓦时单位（1 Wh = 3,600 J）。',
+        `The same measured energy per successful request expressed in watt-hours, a more ` +
+        `familiar household unit (1 Wh = 3,600 J).${MEASURED_TIER_NOTE_EN}`,
+      zh: `与每次成功请求实测能耗相同的量，换算成更直观的瓦时单位（1 Wh = 3,600 J）。${MEASURED_TIER_NOTE_ZH}`,
     },
     formula: {
       en: 'Wh/query = measured J per successful query ÷ 3,600',
@@ -368,11 +382,13 @@ export const METRIC_EXPLANATIONS: Record<MetricKey, MetricExplanation> = {
   measuredPowerPercentTdp: {
     description: {
       en:
-        'Measured average per-chip power as a share of the accelerator’s rated TDP. Values well ' +
-        'below 100% suggest the workload leaves thermal or power headroom on the table.',
+        `Measured average per-chip power as a share of the accelerator’s rated TDP. Values well ` +
+        `below 100% suggest the workload leaves thermal or power headroom on the table.${
+          MEASURED_TIER_NOTE_EN
+        }`,
       zh:
-        '每芯片实测平均功耗占加速器额定 TDP 的百分比。' +
-        '明显低于 100% 说明该工作负载没有用满散热或供电余量。',
+        `每芯片实测平均功耗占加速器额定 TDP 的百分比。` +
+        `明显低于 100% 说明该工作负载没有用满散热或供电余量。${MEASURED_TIER_NOTE_ZH}`,
     },
     formula: {
       en: '% TDP = measured average power per chip (W) ÷ rated TDP (W) × 100',