KernelBench Hard

KernelBench-Hard has moved

Active development now lives in the unified monorepo: **https://github.com/Infatoshi/kernelbench.com/tree/master/benchmarks/hard**

Leaderboard and writeups: **https://kernelbench.com**

This repo is archived (read-only). Its code and full history were merged into the monorepo above.

---

KernelBench-Hard

Surgical GPU kernel benchmark. 6 active CUDA problems, frontier coding agents, roofline-based metric (achieved TFLOPS or GB/s vs hardware peak). Link-don't-spoil problem briefs: agents receive repo/paper URLs, not source snippets.

Sibling project to [KernelBench-v3](https://github.com/Infatoshi/KernelBench-v3) (volume-oriented; local open-weight models). Hard is for frontier-model harnesses on a small, high-signal deck.

PR policy

This repository is published for transparency: it documents the exact prompts, harnesses, traces, kernels, and scoring code I use to evaluate models. It is not an open benchmark track, and I am not accepting PRs that change the problems, hardware target, scoring, prompts, or results. Issues and forks are welcome for discussion or independent experiments, but the canonical repo stays fixed so the published comparisons remain reproducible.

Problem deck

#	Problem	Hardware	What it tests
01	FP8 e4m3 GEMM (off-alignment shapes)	RTX PRO 6000 (SM120)	Tensor-core GEMM, epilogue fusion
02	KDA (Kimi Delta Attention) via CUTLASS CuTe	RTX PRO 6000	Novel attention from paper, CUTLASS 4.x
03	Paged Attention decode	RTX PRO 6000	Indirect indexing, pointer chasing
05	TopK with bitonic sort	RTX PRO 6000	Small-output, comparator networks
06	Sonic-MoE up-projection: grouped GEMM + fused SwiGLU	RTX PRO 6000	Megakernel, load balancing, variable-length
07	W4A16 weight-only GEMM (AWQ/GPTQ-style)	RTX PRO 6000	Bit unpack, quantization, memory-bound decode
08	~~Lightning Attention step (decode) — M4 Max,