Skip to content

Perf Tracker 2026 11 07

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

Perf Tracker 2026-11-07

Nightly run 42351460931, commit 8e1c0eeba, 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 14 of 206
Rows behind 27 of 2333
By label P0 0, P1 0, P2 2, P3 1, P4 11
Failed rows 0

Since nightly run 42257687830:

  • Newly behind: RsqrtFwdOp
  • Caught up: BitwiseXorFwdOp, DivFwdOp, EqFwdOp, GEMMINT8W8A8FwdOp, LtFwdOp, ReciprocalFwdOp
  • 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 GQAVarlenFwdOp attention P2 0.8137 2 of 63 4
2 GEMMFwdOp gemm P2 0.8701 5 of 27 4
3 GroupedGEMMFwdOp gemm P3 0.9284 3 of 10 3
4 RMSNormFwdOp norm P4 0.9661 2 of 34 5
5 SoftmaxFwdOp reduction P4 0.9702 4 of 42 4
6 MoEGroupedGEMMFP8FwdOp moe P4 0.9781 1 of 18 3
7 NeFwdOp elementwise P4 0.9781 2 of 13 3
8 LogicalOrFwdOp elementwise P4 0.9785 2 of 18 3
9 BMMFwdOp gemm P4 0.9827 1 of 16 2
10 GEMMFP8FwdOp gemm P4 0.9847 1 of 39 4
11 GtFwdOp elementwise P4 0.9853 1 of 13 2
12 MaxPool2dIndicesFwdOp pool P4 0.9875 1 of 9 2
13 RemainderFwdOp elementwise P4 0.9886 1 of 14 4
14 RsqrtFwdOp elementwise P4 0.9896 1 of 5 1

Rows behind

GQAVarlenFwdOp

Benchmark benchmarks.ops.bench_gqa.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
llama-8b-chunk-decodes-bfloat16 262.033 fa3 213.217 0.8137
llama-8b-chunk-decodes-float16 264.944 fa3 215.616 0.8138

GEMMFwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen2.5-7b-qkv-b64-bfloat16 13.185 cublaslt-best 11.472 0.8701
ds-v3-m2-qkv-a-bfloat16 12.480 cublaslt-best 12.064 0.9667
ds-v3-m96-qkv-a-bfloat16 14.753 cublaslt-best 14.288 0.9685
qwen3-8b-gate-up-b32-bfloat16 49.120 cublaslt-best 48.032 0.9778
ds-v3-m64-down-bfloat16 9.984 cublaslt-best 9.856 0.9872

GroupedGEMMFwdOp

Benchmark benchmarks.ops.bench_grouped_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-down-bfloat16 48.034 quack 44.593 0.9284
qwen3-30b-ep8-skew-up-bfloat16 83.936 quack 80.609 0.9604
qwen3-30b-ep8-routed1k-bfloat16 31.056 quack 30.672 0.9876

RMSNormFwdOp

Benchmark benchmarks.ops.bench_norm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-8192-float16 256.416 torch-compile 247.713 0.9661
quack-throughput-8192-bfloat16 255.361 torch-compile 251.489 0.9848

SoftmaxFwdOp

Benchmark benchmarks.ops.bench_softmax.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
quack-throughput-512-bfloat16 16.672 quack 16.176 0.9702
quack-throughput-4096-bfloat16 127.840 quack 126.385 0.9886
quack-throughput-2048-bfloat16 63.776 quack 63.120 0.9897
quack-throughput-1024-bfloat16 31.136 quack 30.816 0.9897

MoEGroupedGEMMFP8FwdOp

Benchmark benchmarks.ops.bench_moe_grouped_gemm_fp8.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
qwen3-30b-ep8-skew-up 48.928 deepgemm 47.857 0.9781

NeFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
linear-bias-bfloat16 4.384 torch-compile 4.288 0.9781
linear-bias-float16 4.384 torch-compile 4.288 0.9781

LogicalOrFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
linear-bias-float16 4.416 torch-compile 4.321 0.9785
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-bfloat16 11.072 quack 10.881 0.9827

GEMMFP8FwdOp

Benchmark benchmarks.ops.bench_gemm.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
ds-v3-o-proj-m128-s128-float8_e4m3fn-bfloat16 41.920 flashinfer-fp8-blockscale-sm90 41.280 0.9847

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

MaxPool2dIndicesFwdOp

Benchmark benchmarks.ops.bench_pool.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
vgg-block-float16 9.024 torch-compile 8.911 0.9875

RemainderFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
cnn-feat-broadcast-float32 23.889 torch-compile 23.616 0.9886

RsqrtFwdOp

Benchmark benchmarks.ops.bench_elementwise.

Workload TileOPs (us) Fastest comparator Comparator (us) Ratio
16M-float32 30.752 torch-compile 30.432 0.9896

Clone this wiki locally