cakebench / README.md
Bobliu's picture
Add card and recorded scores
37fb794 verified
|
Raw History Blame Contribute Delete
1.3 kB
metadata
license: mit
pretty_name: cakebench
tags:
  - gpu-kernels
  - benchmark
  - triton
  - cuda

cakebench

A kernel benchmark modelled on CAKE (arXiv 2608.12629). It scores a candidate kernel on 5 kernel families (21 shapes in the full GPU portfolio) by first checking correctness against an fp32 oracle on every shape and 3 input distributions (normal / heavy-tailed / tiny), then timing it (median GPU span, L2 flushed) against a torch baseline.

No model weights and no LLM-generated kernels are released here. This repo holds the recorded scores of the five hand-written example kernels (plain torch) on one shared A100-SXM4-40GB, so you can compare your own harness runs.

task correct shapes geomean vs torch baseline
kmeans_assign 3/3 1.69x
knn_search 3/3 1.13x
tinygemm 8/8 1.00x
kda_decode 4/4 1.03x
kda_prefill 3/3 1.06x

Shared box, noisy timings; the Cake compiler is not public, so this benches kernels, not the compiler; baselines are torch, not the paper's FlashML/FlashKDA/FlashInfer.