Skip to content

Perf Tracker 2026 11 09

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

Perf Tracker 2026-11-09

Nightly run 42502722407, commit ba31dd42f, 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 9 of 206
Rows behind 12 of 2337
By label P0 0, P1 0, P2 0, P3 1, P4 8
Failed rows 0

Since nightly run 42394161005:

  • Newly behind: BMMFwdOp, DivFwdOp, ErfFwdOp
  • Caught up: GtFwdOp, MaxPool2dIndicesFwdOp, MoEGroupedGEMMFP8FwdOp, NeFwdOp, ReciprocalFwdOp, 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 DivFwdOp elementwise P3 0.9232 1 of 20 1
2 GEMMFwdOp gemm P4 0.9654 2 of 27 6
3 GroupedGEMMFwdOp gemm P4 0.9778 1 of 10 5
4 SoftmaxFwdOp reduction P4 0.9786 2 of 42 6
5 RMSNormFwdOp norm P4 0.9818 1 of 34 7
6 ErfFwdOp elementwise P4 0.9849 1 of 5 1
7 LogicalOrFwdOp elementwise P4 0.9853 2 of 18 5
8 BMMFwdOp gemm P4 0.9857 1 of 16 1
9 RemainderFwdOp elementwise P4 0.9866 1 of 14 6

Rows behind

DivFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ragged-chan-broadcast-float16 15.008 torch-compile 13.856 0.9232

GEMMFwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ds-v3-m96-qkv-a-bfloat16 14.784 torch-cublas 14.272 0.9654
qwen3-8b-gate-up-b32-bfloat16 49.072 cublaslt-best 48.096 0.9801

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-up-bfloat16 82.240 quack 80.417 0.9778

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-512-bfloat16 16.480 quack 16.128 0.9786
quack-throughput-2048-bfloat16 63.584 quack 62.704 0.9862

RMSNormFwdOp

Benchmark benchmarks.ops.bench_norm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-256-float16 8.768 torch-compile 8.608 0.9818

ErfFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 31.776 torch 31.296 0.9849

LogicalOrFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
linear-bias-float16 4.416 torch-compile 4.351 0.9853
linear-bias-bfloat16 4.416 torch-compile 4.352 0.9855

BMMFwdOp

Benchmark benchmarks.ops.bench_bmm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
square-512-float16 11.168 quack 11.008 0.9857

RemainderFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-broadcast-float32 23.872 torch-compile 23.553 0.9866

Clone this wiki locally