+
+
+
+ + {text.within}: {panel.stats.withinDieGroupMean?.toFixed(2) ?? text.noPairs} · {text.between} + : {panel.stats.betweenDieGroupMean?.toFixed(2) ?? text.noPairs} SYS_CNT ticks +
+{t.subtitle}
+{t.limits}
+{t.units}
+{t.sharedScale}
+{t.method}
+{t.concurrentLimit}
+ + {t.source} + +
{text.within}: {panel.stats.withinDieGroupMean?.toFixed(2) ?? text.noPairs} · {text.between}
: {panel.stats.betweenDieGroupMean?.toFixed(2) ?? text.noPairs} SYS_CNT ticks
diff --git a/packages/skills/skills/inferencex-api/integrity.json b/packages/skills/skills/inferencex-api/integrity.json
index 353016035..6bbf7dd79 100644
--- a/packages/skills/skills/inferencex-api/integrity.json
+++ b/packages/skills/skills/inferencex-api/integrity.json
@@ -9,7 +9,7 @@
"references/cli-contract.md": "4f070877eeefffad2860c58f79f68432fbe27aa21fc578a1c696f0309a3d7d7b",
"references/cli.md": "11e1183f7ed70d669019e8435ac16a6a9f873de027507b8850a9943f6659bfd9",
"references/collectivex.md": "f79cf8e2779a9289ab69f79f4a4b338e0b9a5f76a3750a3e969f33a744e726a1",
- "references/dashboard-views.md": "1b8a0d95588d173b47aa3de1cde80700370cf82b66b73a1780bc0a1d0c2df86e",
+ "references/dashboard-views.md": "3e910feda374e564211bb5fdcb9660d98fa8e6eafcd0d7317c1b819b6f95baef",
"references/offline-exports.md": "95aa565dcd4c9e592159baa560a9a58217bc61e395de1c6daf0576795c7f4ec9",
"references/pareto.md": "1b4d2d163f982e3f2d97addce310789eae9c1e5badd82dd7501708c9e0385027",
"references/powerx.md": "5af5527d54a7e6191a5769080c122638b333b1ceb59e19fdee7df662a5cd3c8f",
diff --git a/packages/skills/skills/inferencex-api/references/dashboard-views.md b/packages/skills/skills/inferencex-api/references/dashboard-views.md
index acb9313fc..4d66e3c6c 100644
--- a/packages/skills/skills/inferencex-api/references/dashboard-views.md
+++ b/packages/skills/skills/inferencex-api/references/dashboard-views.md
@@ -172,29 +172,29 @@ All paths below are relative to `/api/v1/views/`. The maintained exhaustive
query-key list is in the app's OpenAPI contract. The following groups explain
which API selectors belong together.
-| View | Selection and calculation |
-| ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `options` | Static registries and defaults, JSON only. Data-dependent run/config choices come from their own views. |
-| `inference` | Model, sequence, precisions, date/exact run, GPU/vendor/framework/deployment/spec/power filters; metric, xmode, xstat, xmetric, percentile, optimal/best/allPoints; API-only equal-service sources/target; roleShare and powerFit panels; TCO, custom costs/powers, pricing, unofficial runs, comparison dates/endpoints. JSON or CSV. |
-| `historical` | Model, sequence, target, metric, precisions, GPUs/vendors/frameworks/deployment, start/end, TCO and pricing. `extendToDate` labels synthetic extension, defaults to current UTC date. JSON or CSV. |
-| `calculator` | Model/sequence/run/date, precisions/GPUs, percentile, target/mode, token type, owning/renting cost, TCO, power budget, cost cap, hide-above-limit and public unofficial runs. JSON or CSV. |
-| `first-token` | Benchmark selection plus 1–8 positive TTFT caps in seconds, minimum interactivity, cost provider and token type. Shared dashboard winner selection. |
-| `cache-reuse` | AgentX selection plus exact `config` from returned configurations and `recipe` from data.recipes. Cache-reuse curves retain official versus unofficial evidence. |
-| `profit-estimator` | AgentX selection, comparison dates/runs, target, utilization, lab revenue share, list/OpenRouter/custom token prices, own/rent/custom chip costs, TCO, actual/theoretical cache-hit mode and provisioned/modeled/compare power. USD/chip-hour. |
-| `profit-estimator-per-gigawatt` | Same selectors; USD/GW-year basis and owning-cost default. |
-| `fleet` | Model/sequence/precision/GPUs, target/percentile, TCO/token type/cost provider, MW, input/output prices, ramp, cache discount, MTBI, recovery, horizon and metric. JSON or CSV. |
-| `evaluation` | Model, task, date, precision and GPU selection; public unofficial runs remain separately labeled and independent of the official date cutoff. JSON or CSV. |
-| `reliability` | Rolling range, GPU selection and optional `asOf` date for reproducibility. JSON or CSV. |
-| `gpu-specs` | Full hardware properties or selected metric ranking; table/radar/bar are renderings of these values. JSON or CSV. |
-| `overview` | Models, model/hardware row limits, hardware tier, engine, comparison mode and reference. JSON or CSV. |
-| `rankings` | Ranking kind (model or chip), selected model/scenario and format. Use the exact values in OpenAPI. |
-| `compare` | GPU pairs, model/slug, scenario, tier selection and variant. JSON or CSV. |
-| `collectivex` | Ordered run selection and suite; EP size/phase/modes/precision/operation/percentile/axis/SKU/backend/series; KV page size/x/y/pull/push/overlap ISL/series; swap direction/layout/metric/percentile/series. |
-| `submissions` | Search, table sort/direction/offset/limit, weekly/cumulative chart, on-change cutoff and NVIDIA/AMD/total lines. Table search does not filter the independent submission-volume chart. |
-| `current-inferencex-image` | Model, sequence, hardware, precision, speculation, node type, framework families, `asOf` for image age/release status, and `includeHistory` for daily age history. |
-| `gpu-metrics` | Required run, file/host artifact, GPU indices, metric or correlation axes, statistics sorting, chart mode and interactive downsampling preference. One telemetry source per run; full-record statistics computed on read. Not a serving-window/role-pool projection. Responses are no-store. |
-| `ubenchx` | Test selector (`test`: `mem-bw`, `sm-l2-distance`, `tpc-skyline`, `core-memory-profiles`, or `all`) plus GPU selector (`all` or one GPU). `mem-bw` returns static device-memory copy bandwidth rows with derived MBU and latency. `sm-l2-distance` returns SM-to-SM L2 latency difference matrix with GPC/die topology and summary statistics. `tpc-skyline` returns measured TPCs per GPC (sorted descending) with GPC, TPC, and SM counts, defaulting to every GPU. `core-memory-profiles` returns all four Ascend 950PR_9579 scalar-controller matrices (Cube/Vector, isolated/concurrent), actual core IDs, shared color limits, per-die-group means and provenance. Values are mean absolute differences between matched-address profiles in raw `SYS_CNT ticks`, not core-to-core latency or calibrated cycles. No grid-wide barrier is used in concurrent launches. `all` (default) returns every test. Explicit device filters leave unavailable test sections empty. The Y-axis selector on `/ubenchx/mem-bw` is presentation-only. |
-| `video` | CI run/artifact discovery; only already-published artifact reads; source, serving cell, media/fidelity slot, power phase and GPU denominator; same-workload comparisons, axes, deployment costs and selected tradeoff point. No-store. |
+| View | Selection and calculation |
+| ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| `options` | Static registries and defaults, JSON only. Data-dependent run/config choices come from their own views. |
+| `inference` | Model, sequence, precisions, date/exact run, GPU/vendor/framework/deployment/spec/power filters; metric, xmode, xstat, xmetric, percentile, optimal/best/allPoints; API-only equal-service sources/target; roleShare and powerFit panels; TCO, custom costs/powers, pricing, unofficial runs, comparison dates/endpoints. JSON or CSV. |
+| `historical` | Model, sequence, target, metric, precisions, GPUs/vendors/frameworks/deployment, start/end, TCO and pricing. `extendToDate` labels synthetic extension, defaults to current UTC date. JSON or CSV. |
+| `calculator` | Model/sequence/run/date, precisions/GPUs, percentile, target/mode, token type, owning/renting cost, TCO, power budget, cost cap, hide-above-limit and public unofficial runs. JSON or CSV. |
+| `first-token` | Benchmark selection plus 1–8 positive TTFT caps in seconds, minimum interactivity, cost provider and token type. Shared dashboard winner selection. |
+| `cache-reuse` | AgentX selection plus exact `config` from returned configurations and `recipe` from data.recipes. Cache-reuse curves retain official versus unofficial evidence. |
+| `profit-estimator` | AgentX selection, comparison dates/runs, target, utilization, lab revenue share, list/OpenRouter/custom token prices, own/rent/custom chip costs, TCO, actual/theoretical cache-hit mode and provisioned/modeled/compare power. USD/chip-hour. |
+| `profit-estimator-per-gigawatt` | Same selectors; USD/GW-year basis and owning-cost default. |
+| `fleet` | Model/sequence/precision/GPUs, target/percentile, TCO/token type/cost provider, MW, input/output prices, ramp, cache discount, MTBI, recovery, horizon and metric. JSON or CSV. |
+| `evaluation` | Model, task, date, precision and GPU selection; public unofficial runs remain separately labeled and independent of the official date cutoff. JSON or CSV. |
+| `reliability` | Rolling range, GPU selection and optional `asOf` date for reproducibility. JSON or CSV. |
+| `gpu-specs` | Full hardware properties or selected metric ranking; table/radar/bar are renderings of these values. JSON or CSV. |
+| `overview` | Models, model/hardware row limits, hardware tier, engine, comparison mode and reference. JSON or CSV. |
+| `rankings` | Ranking kind (model or chip), selected model/scenario and format. Use the exact values in OpenAPI. |
+| `compare` | GPU pairs, model/slug, scenario, tier selection and variant. JSON or CSV. |
+| `collectivex` | Ordered run selection and suite; EP size/phase/modes/precision/operation/percentile/axis/SKU/backend/series; KV page size/x/y/pull/push/overlap ISL/series; swap direction/layout/metric/percentile/series. |
+| `submissions` | Search, table sort/direction/offset/limit, weekly/cumulative chart, on-change cutoff and NVIDIA/AMD/total lines. Table search does not filter the independent submission-volume chart. |
+| `current-inferencex-image` | Model, sequence, hardware, precision, speculation, node type, framework families, `asOf` for image age/release status, and `includeHistory` for daily age history. |
+| `gpu-metrics` | Required run, file/host artifact, GPU indices, metric or correlation axes, statistics sorting, chart mode and interactive downsampling preference. One telemetry source per run; full-record statistics computed on read. Not a serving-window/role-pool projection. Responses are no-store. |
+| `ubenchx` | Test selector (`test`: `mem-bw`, `sm-l2-distance`, `tpc-skyline`, `core-memory-profiles`, or `all`) plus GPU selector (`all` or one GPU). `mem-bw` returns static device-memory copy bandwidth rows with derived MBU and latency. `sm-l2-distance` returns SM-to-SM L2 latency difference matrix with GPC/die topology and summary statistics. `tpc-skyline` returns measured TPCs per GPC (sorted descending) with GPC, TPC, and SM counts, defaulting to every GPU. `core-memory-profiles` returns all four Ascend 950PR_9579 scalar-controller matrices (Cube/Vector, isolated/concurrent), actual core IDs, shared color limits, per-die-group means and provenance. Values are mean absolute differences between matched-address profiles in raw `SYS_CNT ticks`, not core-to-core latency or calibrated cycles. No grid-wide barrier is used in concurrent launches. Cell hover, pinning and clear-selection controls are presentation-only; all matrix cells remain available in the API. `all` (default) returns every test. Explicit device filters leave unavailable test sections empty. The Y-axis selector on `/ubenchx/mem-bw` is presentation-only. |
+| `video` | CI run/artifact discovery; only already-published artifact reads; source, serving cell, media/fidelity slot, power phase and GPU denominator; same-workload comparisons, axes, deployment costs and selected tradeoff point. No-store. |
## Interpretation and maintenance
From bd805dc498686c112a8d81a82b7d8d12acb47c73 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 26 Oct 2026 23:40:15 +0800
Subject: [PATCH 4/4] feat(ubenchx): combine B200 and Huawei core profiles
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
中文：将 B200 与 Huawei 核心访存特征合并到同一页面，提供硬件筛选，保留独立单位与色标，并兼容旧页面链接和 API。
---
docs/ubenchx-ascend-core-profiles.md | 10 ++-
packages/app/cypress/e2e/ubenchx.cy.ts | 46 ++++++++++--
.../app/api/v1/views/ubenchx/route.test.ts | 29 +++++++-
.../app/src/app/api/v1/views/ubenchx/route.ts | 15 ++--
.../app/ubenchx/core-memory-profiles/page.tsx | 2 +-
.../src/app/ubenchx/sm-l2-distance/page.tsx | 37 +---------
.../zh/ubenchx/core-memory-profiles/page.tsx | 2 +-
.../app/zh/ubenchx/sm-l2-distance/page.tsx | 38 +---------
.../ubenchx/CoreMemoryProfilesContent.tsx | 71 ++++++++++++++++++-
.../src/components/ubenchx/SmL2Heatmap.tsx | 32 ++++-----
.../app/src/components/ubenchx/UbenchxHub.tsx | 3 -
.../ubenchx/core-memory-profiles.test.ts | 10 ++-
.../ubenchx/core-memory-profiles.ts | 24 +++++++
packages/app/src/lib/api/api-route-catalog.ts | 4 +-
.../app/src/lib/views-api/docs/extensions.ts | 12 ++--
.../skills/inferencex-api/integrity.json | 2 +-
.../references/dashboard-views.md | 48 +++++++------
17 files changed, 239 insertions(+), 146 deletions(-)
diff --git a/docs/ubenchx-ascend-core-profiles.md b/docs/ubenchx-ascend-core-profiles.md
index 0df8f2c85..f80f0f0c1 100644
--- a/docs/ubenchx-ascend-core-profiles.md
+++ b/docs/ubenchx-ascend-core-profiles.md
@@ -1,6 +1,10 @@
-# Ascend scalar-controller memory profiles
+# Core memory-access profiles: B200 and Huawei 950PR
-The `core-memory-profiles` view preserves the four 128 MiB matrices from the 6–7 October 2026 (Singapore time) Cube/Vector locality rerun: 28 Cube scalar controllers and 56 Vector scalar controllers, each tested with one active core and with concurrent launches. Both core types use identical scalar pointer-chase instructions on the same allocation. Each address is averaged over six passes before computing the mean absolute difference between two cores' address profiles.
+The `/ubenchx/core-memory-profiles` page combines B200 and Huawei 950PR, showing both by default. Its hardware selector can focus on either device. The B200 SM matrix keeps its `clock64` cycle units, GPC/die grouping, statistics and independent color domain; the 950PR retains SYS_CNT ticks and a shared scale across its four panels. Colors and values across devices are not comparable absolute latency. Both matrices measure differences between matched-address latency profiles, not direct core-to-core communication time.
+
+The legacy `/ubenchx/sm-l2-distance` page and its Chinese sibling redirect to the combined page; the existing `test=sm-l2-distance` API remains compatible.
+
+The Ascend portion of the `core-memory-profiles` view preserves the four 128 MiB matrices from the 6–7 October 2026 (Singapore time) Cube/Vector locality rerun: 28 Cube scalar controllers and 56 Vector scalar controllers, each tested with one active core and with concurrent launches. Both core types use identical scalar pointer-chase instructions on the same allocation. Each address is averaged over six passes before computing the mean absolute difference between two cores' address profiles.
The data file contains every displayed matrix entry without rounding, the actual physical core-ID lists, the original pair statistics and source hashes. The arrays preserve ID gaps. Runtime-ID die groups divide IDs below 54 from those at or above 54; they do not assign physical coordinates. The frontend and API recompute within/between-group means from unique off-diagonal pairs using the same pure function. A shared 0–113.63708496093749 SYS_CNT-tick color domain applies to all four panels.
@@ -8,4 +12,4 @@ The unit is raw `SYS_CNT ticks`, not `clock64` cycles or nanoseconds. These are
The complete displayed data is in `ascend-core-profiles.json`. Its `sourceSha256` identifies the original plotted-data JSON; each panel also retains the hashes of its archived analysis inputs. This publication contains the displayed matrices and aggregate statistics, not every original per-load observation.
-Coverage: `/ubenchx/core-memory-profiles` and its `/zh` sibling show all four panels. Hover, keyboard arrows and pinning only inspect already-returned cells; they do not change data. A repeated click or Enter unpins the current cell. Clicking outside the panel, Escape, or Clear selection clears it. A click on a different cell changes the pin; pointer exit clears an unpinned hover preview. The hub's new test is available through `/api/v1/views/ubenchx?test=core-memory-profiles`, with `gpu=all` or `gpu=Ascend 950PR_9579`, and as `coreMemoryProfiles` in `test=all`. Explicit device selections leave other unavailable test sections empty. API `unit`, `metric`, `limitations`, `coreIds`, `dieGroups` and `colorLimits` preserve the measurement interpretation.
+Coverage: `/ubenchx/core-memory-profiles` and its `/zh` sibling show B200 and all four Ascend panels. Hover, keyboard arrows and pinning only inspect already-returned cells; they do not change data. A repeated click or Enter unpins the current cell. Clicking outside the panel, Escape, or Clear selection clears it. A click on a different cell changes the pin; pointer exit clears an unpinned hover preview. The hub's new test is available through `/api/v1/views/ubenchx?test=core-memory-profiles`, with `gpu=all` (default), `gpu=B200 SXM` or `gpu=Ascend 950PR_9579`, and as `coreMemoryProfiles` in `test=all`. Explicit device selections leave other unavailable test sections empty. B200 keeps its original transformed SM fields plus `unit`, `metric`, `limitations` and `colorLimits`; Ascend keeps its panel schema. API `unit`, `metric`, `limitations`, `coreIds`, `dieGroups` and `colorLimits` preserve the measurement interpretation.
diff --git a/packages/app/cypress/e2e/ubenchx.cy.ts b/packages/app/cypress/e2e/ubenchx.cy.ts
index 5637e2452..d5dd21703 100644
--- a/packages/app/cypress/e2e/ubenchx.cy.ts
+++ b/packages/app/cypress/e2e/ubenchx.cy.ts
@@ -26,7 +26,7 @@ describe('ubenchX hub and test routes', () => {
cy.get('[data-testid="ubenchx-metric-select"]').should('contain', 'TB/s');
});
- it('switches to sm-l2-distance with the test selector', () => {
+ it('switches to the combined core profiles with the test selector', () => {
cy.visit('/ubenchx/mem-bw', dismissModal);
cy.contains('ubenchX').should('be.visible');
cy.get('[data-testid="ubenchx-metric-select"]').should('be.visible');
@@ -36,10 +36,10 @@ describe('ubenchX hub and test routes', () => {
.should('include', '/zh/ubenchx/mem-bw');
cy.get('[data-testid="ubenchx-test-select"]').click();
- cy.contains('[role="option"]', 'SM-SM L2 Latency Difference').click();
- cy.url().should('include', '/ubenchx/sm-l2-distance');
+ cy.contains('[role="option"]', 'Core memory-access profiles').click();
+ cy.url().should('include', '/ubenchx/core-memory-profiles');
cy.get('[data-testid="sm-l2-heatmap"]').should('be.visible');
- cy.get('[data-testid="ubenchx-test-select"]').should('contain', 'SM-SM L2 Latency Difference');
+ cy.get('[data-testid="ubenchx-test-select"]').should('contain', 'Core memory-access profiles');
});
it('switches to the TPC Skyline table with the test selector', () => {
@@ -90,7 +90,7 @@ describe('ubenchX hub and test routes', () => {
cy.get('meta[name="robots"]').should('have.attr', 'content').and('contain', 'noindex');
cy.get('link[rel="alternate"][hreflang="zh-CN"]')
.invoke('attr', 'href')
- .should('include', '/zh/ubenchx/sm-l2-distance');
+ .should('include', '/zh/ubenchx/core-memory-profiles');
cy.get('[data-testid="sm-l2-heatmap"] svg rect').last().as('overlay');
cy.get('@overlay').click(520, 120, { force: true });
@@ -195,6 +195,42 @@ it('unpins core-profile cells by repeated click or Enter and clears them outside
cy.get(clear).should('be.disabled');
});
+it('filters both devices on one page and preserves their API values and units', () => {
+ cy.visit('/ubenchx/core-memory-profiles', dismiss);
+ cy.get('[data-testid="sm-l2-heatmap"]').should('exist');
+ cy.get('[data-testid="ascend-core-profiles"]').should('exist');
+ cy.get('[data-testid="core-profile-hardware"]').should('contain', 'All hardware').click();
+ cy.contains('[role="option"]', 'B200 SXM').click();
+ cy.get('[data-testid="sm-l2-heatmap"]').should('be.visible');
+ cy.get('[data-testid="ascend-core-profiles"]').should('not.exist');
+ cy.request('/api/v1/views/ubenchx?test=core-memory-profiles&gpu=B200%20SXM').then(({ body }) => {
+ expect(body.gpus).to.have.length(1);
+ expect(body.gpus[0].unit).to.equal('clock64 cycles');
+ cy.contains(`${body.gpus[0].stats.crossDieMean.toFixed(1)} cycles`).should('be.visible');
+ });
+ cy.get('[data-testid="core-profile-hardware"]').click();
+ cy.contains('[role="option"]', 'Huawei 950PR').click();
+ cy.get('[data-testid="sm-l2-heatmap"]').should('not.exist');
+ cy.get('[data-testid="profile-scalar_128MiB_cube_isolated"] svg').focus();
+ cy.get('[data-testid="profile-scalar_128MiB_cube_isolated-value"]').should(
+ 'contain',
+ '0 ↔ 1: 19.03 SYS_CNT ticks',
+ );
+ cy.get('[data-testid="core-profile-hardware"]').click();
+ cy.contains('[role="option"]', 'All hardware').click();
+ cy.get('[data-testid="sm-l2-heatmap"]').should('exist');
+ cy.get('[data-testid="ascend-core-profiles"]').should('exist');
+ cy.screenshot('combined-core-profiles-desktop', { capture: 'fullPage' });
+ cy.visit('/zh/ubenchx/sm-l2-distance', dismiss);
+ cy.location('pathname').should('equal', '/zh/ubenchx/core-memory-profiles');
+ cy.viewport(390, 844);
+ cy.get('[data-testid="core-profile-hardware"]').should('contain', '全部硬件');
+ cy.contains('不同设备上颜色或数值相同，并不代表延迟相同').should('be.visible');
+ cy.get('[data-testid="sm-l2-heatmap"]').should('exist');
+ cy.get('[data-testid="ascend-core-profiles"]').should('exist');
+ cy.screenshot('combined-core-profiles-mobile-zh', { capture: 'fullPage' });
+});
+
describe('Huawei copy-bandwidth evidence', () => {
it('renders Huawei 950PR and qualifies the graph timing in English and Chinese', () => {
cy.visit('/ubenchx/mem-bw', dismiss);
diff --git a/packages/app/src/app/api/v1/views/ubenchx/route.test.ts b/packages/app/src/app/api/v1/views/ubenchx/route.test.ts
index ae1816348..fd72b407d 100644
--- a/packages/app/src/app/api/v1/views/ubenchx/route.test.ts
+++ b/packages/app/src/app/api/v1/views/ubenchx/route.test.ts
@@ -176,13 +176,40 @@ it('returns the Ascend core profiles with raw timer units and leaves unavailable
expect(run.panels[0].stats.withinDieGroupMean).toBeCloseTo(18.55499211685, 8);
expect(run.panels[0].stats.betweenDieGroupMean).toBeCloseTo(93.12987720723, 8);
expect(run.panels[0].matrix[0][14]).toBeCloseTo(96.70648193359, 8);
- const onlyResponse = await GET(request({ test: 'core-memory-profiles' }));
+ const onlyResponse = await GET(
+ request({ test: 'core-memory-profiles', gpu: 'Ascend 950PR_9579' }),
+ );
const only = await onlyResponse.json();
expect(only.gpus).toEqual(body.coreMemoryProfiles.gpus);
const missing = await GET(request({ test: 'core-memory-profiles', gpu: 'H100 SXM' }));
expect(missing.status).toBe(400);
});
+it('combines both hardware profiles without converting or pooling their units', async () => {
+ const response = await GET(request({ test: 'core-memory-profiles' }));
+ const body = await response.json();
+ expect(body.params).toEqual({ test: 'core-memory-profiles', gpu: 'all' });
+ expect(body.gpus.map((run: { gpu: string }) => run.gpu)).toEqual([
+ 'B200 SXM',
+ 'Ascend 950PR_9579',
+ ]);
+ const b200 = body.gpus[0];
+ expect(b200).toMatchObject(transformSmL2Run('B200 SXM', SM_L2_RUNS['B200 SXM']));
+ expect(b200.unit).toBe('clock64 cycles');
+ expect(b200.colorLimits).toEqual([0, b200.stats.max]);
+ expect(body.gpus[1].panels[0].unit).toBe('SYS_CNT ticks');
+ const filteredResponse = await GET(request({ test: 'core-memory-profiles', gpu: 'b200 sxm' }));
+ const filtered = await filteredResponse.json();
+ expect(filtered.params.gpu).toBe('B200 SXM');
+ expect(filtered.gpus).toEqual([b200]);
+ const allResponse = await GET(request({ test: 'all', gpu: 'B200 SXM' }));
+ const all = await allResponse.json();
+ expect(all.coreMemoryProfiles.gpus).toEqual([b200]);
+ const legacyResponse = await GET(request({ test: 'sm-l2-distance', gpu: 'B200 SXM' }));
+ const legacy = await legacyResponse.json();
+ expect(legacy.gpus[0]).toEqual(transformSmL2Run('B200 SXM', SM_L2_RUNS['B200 SXM']));
+});
+
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);
diff --git a/packages/app/src/app/api/v1/views/ubenchx/route.ts b/packages/app/src/app/api/v1/views/ubenchx/route.ts
index 29ef2b8ad..5ad92c5e3 100644
--- a/packages/app/src/app/api/v1/views/ubenchx/route.ts
+++ b/packages/app/src/app/api/v1/views/ubenchx/route.ts
@@ -1,4 +1,8 @@
-import { CORE_PROFILE_GPU, coreMemoryProfiles } from '@/components/ubenchx/core-memory-profiles';
+import {
+ CORE_PROFILE_GPU,
+ CORE_PROFILE_GPU_KEYS,
+ coreMemoryProfileResults,
+} from '@/components/ubenchx/core-memory-profiles';
import { SM_L2_RUNS } from '@/components/ubenchx/sm-l2-data';
import { transformSmL2Run } from '@/components/ubenchx/sm-l2-transform';
import { TPC_SKYLINE_RUNS } from '@/components/ubenchx/tpc-skyline-data';
@@ -35,14 +39,14 @@ export function GET(request: NextRequest) {
const rawGpu = s.get('gpu');
if (test === 'core-memory-profiles') {
- const gpu = parseEnumParam(rawGpu, 'gpu', ['all', CORE_PROFILE_GPU], CORE_PROFILE_GPU);
+ const gpu = parseEnumParam(rawGpu, 'gpu', ['all', ...CORE_PROFILE_GPU_KEYS], 'all');
return Promise.resolve(
cachedJson({
apiVersion: 'v1',
view: 'ubenchx',
test,
params: { test, gpu },
- gpus: [coreMemoryProfiles()],
+ gpus: coreMemoryProfileResults(gpu),
}),
);
}
@@ -105,10 +109,7 @@ export function GET(request: NextRequest) {
memBw: { gpus: memBwResults },
smL2Distance: { gpus: smL2Results },
coreMemoryProfiles: {
- gpus:
- rawGpu === null || gpu === 'all' || gpu === CORE_PROFILE_GPU
- ? [coreMemoryProfiles()]
- : [],
+ gpus: coreMemoryProfileResults(rawGpu === null ? 'all' : gpu),
},
// 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.
diff --git a/packages/app/src/app/ubenchx/core-memory-profiles/page.tsx b/packages/app/src/app/ubenchx/core-memory-profiles/page.tsx
index 76835194f..e60db61ae 100644
--- a/packages/app/src/app/ubenchx/core-memory-profiles/page.tsx
+++ b/packages/app/src/app/ubenchx/core-memory-profiles/page.tsx
@@ -7,7 +7,7 @@ import { SITE_URL } from '@semianalysisai/inferencex-constants';
const TITLE = 'ubenchX: Core memory-access profiles';
const DESCRIPTION =
- 'Measured 128 MiB memory-access profile differences for Ascend 950PR Cube and Vector scalar controllers, with isolated and concurrent launches.';
+ 'Compare B200 SM and Huawei 950PR Cube/Vector memory-access profiles, preserving each device’s native timer units and measurement conditions.';
export const metadata: Metadata = {
title: TITLE,
diff --git a/packages/app/src/app/ubenchx/sm-l2-distance/page.tsx b/packages/app/src/app/ubenchx/sm-l2-distance/page.tsx
index 68d3f3ef4..2a2db9944 100644
--- a/packages/app/src/app/ubenchx/sm-l2-distance/page.tsx
+++ b/packages/app/src/app/ubenchx/sm-l2-distance/page.tsx
@@ -1,36 +1,5 @@
-import type { Metadata } from 'next';
+import { permanentRedirect } from 'next/navigation';
-import { SmL2Content } from '@/components/ubenchx/SmL2Heatmap';
-import { UbenchxHub } from '@/components/ubenchx/UbenchxHub';
-import { enAlternates } from '@/lib/i18n/i18n';
-import { SITE_URL } from '@semianalysisai/inferencex-constants';
-
-const TITLE = 'ubenchX: SM-SM L2 Latency Difference';
-const DESCRIPTION =
- 'L2 pointer-chase microbenchmark showing SM-to-SM latency differences, revealing GPC and die topology structure on NVIDIA Blackwell GPUs.';
-
-export const metadata: Metadata = {
- title: TITLE,
- description: DESCRIPTION,
- alternates: enAlternates('/ubenchx/sm-l2-distance'),
- robots: { index: false, follow: false },
- openGraph: {
- title: `${TITLE} | InferenceX`,
- description: DESCRIPTION,
- url: `${SITE_URL}/ubenchx/sm-l2-distance`,
- },
- twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
-};
-
-export default function UbenchxSmL2DistancePage() {
- return (
-
-
-
-
-
-
-
-
-
{t.subtitle}
{t.limits}
@@ -309,3 +330,47 @@ export function CoreMemoryProfilesContent() {{t.comparison}
+{t.pageSubtitle}
-