kernelbench.com: Agentic GPU Kernel Benchmark Results and Run Artifacts

Arledge, Elliot

GLM-5.2

Z.ai1/9 flagged

One row per published session deck. Score bars are peak fraction of the board roofline (Hard / CUDA) or best speedup vs the torch baseline (Mega). Audit chips come from the human/subagent reward-hack review of every published cell.

hard

zai-claude/glm-5.26/60.851/7 flagged

RTX PRO 6000

FP8 GEMMpass

40.6%clean

session 1h 36m

solution trace

KDA CUTLASSpass

3.2%clean

session 4h 1m

solution trace

Paged Attentionpass

67.7%interesting

session 3h 54m

solution trace

TopK Bitonicpass

3.4%interesting

session 3h 56m

solution trace

Sonic MoEpass

9.8%clean

session 4h 1m

solution trace

W4A16 GEMMpass

32.1%clean

session 4h 40m

solution trace

B200

FP8 GEMMpass

20.0%unaudited

session 1h 40m

solution trace

KDA CUTLASSpass

1.1%unaudited

session 1h 40m

solution trace

Paged Attentionpass

26.4%unaudited

session 1h 41m

solution trace

TopK Bitonicpass

0.2%unaudited

session 45m

solution trace

Sonic MoEpass

5.8%unaudited

session 49m

solution trace

W4A16 GEMMpass

4.8%unaudited

session 1h 45m

solution trace

H100 PCIe

FP8 GEMMpass

7.8%unaudited

session 48m

solution trace

KDA CUTLASSfail

unaudited

session 45m

trace

Paged Attentionpass

31.4%unaudited

session 51m

solution trace

TopK Bitonicpass

0.3%unaudited

session 53m

solution trace

Sonic MoEpass

5.8%unaudited

session 55m

solution trace

W4A16 GEMMpass

8.3%unaudited

session 55m

solution trace

RTX 3090

FP8 GEMMno run

KDA CUTLASSpass

9.7%unaudited

session 1h 19m

trace

Paged Attentionpass

65.8%unaudited

session 29m

trace

TopK Bitonicpass

4.9%unaudited

session 1h 5m

trace

Sonic MoEpass

14.9%unaudited

session 1h 55m

trace

W4A16 GEMMpass

32.3%unaudited

session 2h 0m

trace

rubric leak20260613_053946_zai-claude_glm-5.2_01_fp8_gemm

Rubric leak: a real, well-tuned Triton kernel, but it loads the fp8 activation and immediately upcasts to bf16 in-register (a=tl.load(...).to(tl.bfloat16)) and runs a bf16 tl.dot; the docstring states FP8 quant cannot meet the bf16-reference tolerance. It also sets allow_bf16_reduced_precision_reduction=False to match cuBLAS. Correct but measures bf16, not fp8 tensor-core skill, so it hard-caps near 0.5 peak (0.4925).

mega

zai-claude/glm-5.21/20.602 audited

RTX PRO 6000

RL Grid PPOno run

Kimi-Linear Decodepass

11.14xclean

1951 tok/s 2048 ctx 10.75x8192 ctx 10.94x16384 ctx 11.76x · triton

solution trace

B200

RL Grid PPOno run

Kimi-Linear Decodepass

7.30xclean

815 tok/s 2048 ctx 6.88x8192 ctx 7.25x16384 ctx 7.79x · triton

solution trace

Methodology: Rank per bench: valid passes (audited-clean correct cells / problems) desc, then mean normalized performance (cell score / board best per problem) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.