Skip to content

Perf Tracker 2026 11 12

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

Perf Tracker 2026-11-12

Nightly run 43063959201, 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 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 42860420524:

  • Newly behind: AddFwdOp
  • Caught up: AbsFwdOp, ClampTensorFwdOp, ExpFwdOp, MaskedFillScalarFwdOp, SigmoidFwdOp, WhereFwdOp
  • 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 0.9742 2 of 42 9
2 ErfFwdOp elementwise P4 0.9782 1 of 5 2
3 GroupedGEMMFwdOp gemm P4 0.9807 1 of 10 8
4 AddFwdOp elementwise P4 0.9892 1 of 12 1
5 GEMMW4A16FwdOp gemm P4 0.9897 1 of 15 2

Rows behind

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-32768-float16 1052.840 quack 1025.720 0.9742
quack-throughput-32768-bfloat16 1077.370 quack 1063.690 0.9873

ErfFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 32.255 torch 31.552 0.9782

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-down-bfloat16 42.256 quack 41.440 0.9807

AddFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-broadcast-float32 23.616 torch-compile 23.360 0.9892

GEMMW4A16FwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
chunked-prefill-tile-float16 37.184 torch-dequantized-matmul 36.801 0.9897

Clone this wiki locally