Skip to content

Perf Tracker 2026 11 15

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

Perf Tracker 2026-11-15

Nightly run 43654108754, commit abe2086ef, 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:local-runner-2.338.2. 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.

Summary

Measure Value
Ops behind 5 of 206
Rows behind 6 of 2337
By label P0 0, P1 0, P2 0, P3 0, P4 5
Failed rows 0

Since nightly run 43264767627:

  • Newly behind: FusedMoEExpertsFwdOp, GEMMFwdOp, WhereFwdOp
  • Caught up: ExpFwdOp, FloorFwdOp, RoundFwdOp, SoftmaxFwdOp, TruncFwdOp
  • 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 GroupedGEMMFwdOp gemm P4 0.9765 1 of 10 10
2 ErfFwdOp elementwise P4 0.9811 1 of 5 4
3 GEMMFwdOp gemm P4 0.9866 1 of 27 1
4 FusedMoEExpertsFwdOp moe P4 0.9896 1 of 11 1
5 WhereFwdOp elementwise P4 0.9897 2 of 8 1

Rows behind

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-down-bfloat16 42.241 quack 41.249 0.9765

ErfFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 32.192 torch 31.585 0.9811

GEMMFwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ds-v3-prefill-o-proj-bfloat16 1270.730 quack 1253.650 0.9866

FusedMoEExpertsFwdOp

Benchmark benchmarks.ops.bench_fused_moe_experts.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ds-v3-prefill-bfloat16 7555.090 flashinfer-cutlass 7476.550 0.9896

WhereFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 48.256 torch-compile 47.761 0.9897
attn-causal-mask-float32 62.224 torch-compile 61.601 0.9900

Clone this wiki locally