Skip to content

Perf Tracker 2026 11 11

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

Perf Tracker 2026-11-11

Nightly run 42860420524, commit 9bf4bf56c, 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 10 of 206
Rows behind 11 of 2337
By label P0 0, P1 0, P2 0, P3 0, P4 10
Failed rows 0

Since nightly run 42660464876:

  • Newly behind: AbsFwdOp, ErfFwdOp, ExpFwdOp, GEMMW4A16FwdOp, MaskedFillScalarFwdOp, WhereFwdOp
  • Caught up: BitwiseAndFwdOp, FloorFwdOp, GEMMFwdOp
  • 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 SoftmaxFwdOp reduction P4 0.9744 2 of 42 8
2 GroupedGEMMFwdOp gemm P4 0.9792 1 of 10 7
3 ErfFwdOp elementwise P4 0.9829 1 of 5 1
4 MaskedFillScalarFwdOp elementwise P4 0.9887 1 of 8 1
5 GEMMW4A16FwdOp gemm P4 0.9888 1 of 15 1
6 ClampTensorFwdOp elementwise P4 0.9892 1 of 18 2
7 AbsFwdOp elementwise P4 0.9896 1 of 5 1
8 ExpFwdOp elementwise P4 0.9896 1 of 5 1
9 SigmoidFwdOp elementwise P4 0.9896 1 of 5 2
10 WhereFwdOp elementwise P4 0.9898 1 of 8 1

Rows behind

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-32768-float16 1051.060 quack 1024.180 0.9744
quack-throughput-32768-bfloat16 1078.610 quack 1065.040 0.9874

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-down-bfloat16 42.336 quack 41.456 0.9792

ErfFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 31.872 torch 31.328 0.9829

MaskedFillScalarFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 33.919 torch-compile 33.536 0.9887

GEMMW4A16FwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
prefill-mlp-up-float16 103.857 torch-dequantized-matmul 102.689 0.9888

ClampTensorFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-channel-bounds-float32 23.680 torch-compile 23.425 0.9892

AbsFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.784 torch-compile 30.464 0.9896

ExpFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.880 torch-compile 30.560 0.9896

SigmoidFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.880 torch-compile 30.560 0.9896

WhereFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
attn-causal-mask-float32 62.464 torch-compile 61.824 0.9898

Clone this wiki locally