Skip to content

Perf Tracker 2026 11 16

github-actions[bot] edited this page Nov 15, 2026 · 1 revision

Perf Tracker 2026-11-16

Nightly run 43738645759, commit e4444654b, NVIDIA H200, CUDA 13.2, driver 580.178.04, torch 2.13.0+cu132, TileLang 0.1.12, image ghcr.io/tile-ai/tileops-ci-runner@sha256:5c574e21c3fe62d4da0592cb0d039b5221d37636003bfe82ad4eacd5461be934. Times are CUPTI device_busy_ms.

A row is behind when its fastest comparator's time over TileOPs' time is below 0.99. A row whose fastest comparator takes under 10 us is behind only if it was also below that ratio the night before. An op is labelled by its worst row: P0 < 0.50, P1 < 0.80, P2 < 0.90, P3 < 0.95, P4 < 0.99.

What was measured or compared changed since the previous nightly:

  • image: ghcr.io/tile-ai/tileops-ci-runner:local-runner-2.338.2 -> ghcr.io/tile-ai/tileops-ci-runner@sha256:5c574e21c3fe62d4da0592cb0d039b5221d37636003bfe82ad4eacd5461be934
  • image_digest: None -> sha256:5c574e21c3fe62d4da0592cb0d039b5221d37636003bfe82ad4eacd5461be934

Summary

Measure Value
Ops behind 2 of 206
Rows behind 2 of 2337
By label P0 0, P1 0, P2 0, P3 0, P4 2
Failed rows 0

Since nightly run 43654108754:

  • Newly behind: none
  • Caught up: FusedMoEExpertsFwdOp, GEMMFwdOp, GroupedGEMMFwdOp
  • Not measured tonight: none

Queue

Nightlies behind counts the consecutive nightlies, up to 30, in which the op had a row behind.

# Op Category Label Worst ratio Rows behind Nightlies behind
1 ErfFwdOp elementwise P4 0.9772 1 of 5 5
2 WhereFwdOp elementwise P4 0.9880 1 of 8 2

Rows behind

ErfFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 32.224 torch 31.488 0.9772

WhereFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
attn-causal-mask-float32 62.496 torch-compile 61.745 0.9880

Clone this wiki locally