Skip to content

Perf Tracker 2026 11 10

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

Perf Tracker 2026-11-10

Nightly run 42660464876, commit 368c2dcde, 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 7 of 206
Rows behind 8 of 2337
By label P0 0, P1 0, P2 0, P3 0, P4 7
Failed rows 0

Since nightly run 42502722407:

  • Newly behind: BitwiseAndFwdOp, ClampTensorFwdOp, FloorFwdOp, SigmoidFwdOp
  • Caught up: BMMFwdOp, DivFwdOp, ErfFwdOp, LogicalOrFwdOp, RMSNormFwdOp, RemainderFwdOp
  • 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 0.9745 2 of 42 7
2 GroupedGEMMFwdOp gemm P4 0.9781 1 of 10 6
3 GEMMFwdOp gemm P4 0.9841 1 of 27 7
4 ClampTensorFwdOp elementwise P4 0.9853 1 of 18 1
5 FloorFwdOp elementwise P4 0.9896 1 of 5 1
6 SigmoidFwdOp elementwise P4 0.9896 1 of 5 1
7 BitwiseAndFwdOp elementwise P4 0.9898 1 of 14 1

Rows behind

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-32768-float16 1052.770 quack 1025.910 0.9745
quack-throughput-32768-bfloat16 1079.640 quack 1065.160 0.9866

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-down-bfloat16 42.304 quack 41.376 0.9781

GEMMFwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ds-v3-prefill-o-proj-bfloat16 1272.300 deepgemm 1252.110 0.9841

ClampTensorFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-channel-bounds-float32 23.839 torch-compile 23.488 0.9853

FloorFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.591 torch-compile 30.272 0.9896

SigmoidFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.815 torch-compile 30.495 0.9896

BitwiseAndFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-broadcast-int32 23.569 torch-compile 23.328 0.9898

Clone this wiki locally