Skip to content

Perf Tracker 2026 11 08

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

Perf Tracker 2026-11-08

Nightly run 42394161005, commit 31c92ddaf, 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 12 of 206
Rows behind 25 of 2333
By label P0 0, P1 0, P2 2, P3 0, P4 10
Failed rows 0

Since nightly run 42351460931:

  • Newly behind: ReciprocalFwdOp, WhereFwdOp
  • Caught up: BMMFwdOp, GEMMFP8FwdOp, GQAVarlenFwdOp, RsqrtFwdOp
  • 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 P2 0.8587 3 of 10 4
2 GEMMFwdOp gemm P2 0.8710 6 of 27 5
3 RMSNormFwdOp norm P4 0.9660 4 of 34 6
4 MoEGroupedGEMMFP8FwdOp moe P4 0.9731 1 of 18 4
5 NeFwdOp elementwise P4 0.9782 2 of 13 4
6 SoftmaxFwdOp reduction P4 0.9787 2 of 42 5
7 MaxPool2dIndicesFwdOp pool P4 0.9824 1 of 9 3
8 GtFwdOp elementwise P4 0.9853 1 of 13 3
9 LogicalOrFwdOp elementwise P4 0.9855 2 of 18 4
10 RemainderFwdOp elementwise P4 0.9860 1 of 14 5
11 ReciprocalFwdOp elementwise P4 0.9886 1 of 5 1
12 WhereFwdOp elementwise P4 0.9887 1 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 48.145 quack 41.344 0.8587
qwen3-30b-ep8-skew-up-bfloat16 83.984 quack 80.031 0.9529
qwen3-30b-ep8-routed1k-bfloat16 30.977 quack 30.592 0.9876

GEMMFwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen2.5-7b-qkv-b64-bfloat16 13.152 cublaslt-best 11.456 0.8710
ds-v3-m96-qkv-a-bfloat16 14.816 cublaslt-best 14.304 0.9654
ds-v3-m2-qkv-a-bfloat16 12.416 cublaslt-best 12.002 0.9666
qwen3-8b-gate-up-b32-bfloat16 49.312 cublaslt-best 48.128 0.9760
ds-v3-m64-down-bfloat16 9.984 cublaslt-best 9.856 0.9872
ds-v3-prefill-down-bfloat16 170.321 cublaslt-best 168.337 0.9884

RMSNormFwdOp

Benchmark benchmarks.ops.bench_norm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-8192-float16 256.193 torch-compile 247.473 0.9660
quack-throughput-8192-bfloat16 255.633 torch-compile 251.089 0.9822
quack-throughput-256-float16 8.800 torch-compile 8.672 0.9855
quack-throughput-256-bfloat16 8.800 torch-compile 8.673 0.9856

MoEGroupedGEMMFP8FwdOp

Benchmark benchmarks.ops.bench_moe_grouped_gemm_fp8.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-up 48.832 deepgemm 47.520 0.9731

NeFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
linear-bias-bfloat16 4.384 torch-compile 4.288 0.9782
linear-bias-float16 4.384 torch-compile 4.320 0.9854

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-512-bfloat16 16.512 quack 16.160 0.9787
quack-throughput-2048-bfloat16 63.680 quack 62.688 0.9844

MaxPool2dIndicesFwdOp

Benchmark benchmarks.ops.bench_pool.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
vgg-block-float16 9.088 torch-compile 8.928 0.9824

GtFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
linear-bias-bfloat16 4.352 torch-compile 4.288 0.9853

LogicalOrFwdOp

Benchmark benchmarks.ops.bench_elementwise.

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

RemainderFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-broadcast-float32 23.888 torch-compile 23.553 0.9860

ReciprocalFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.784 torch-compile 30.432 0.9886

WhereFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
attn-causal-mask-float32 62.416 torch-compile 61.713 0.9887

Clone this wiki locally