Kernelbench.Com

KernelBench

Frontier coding agents write GPU kernels. Each session is one autonomous agent, graded against a roofline (or ms/speedup) ceiling, then reward-hack audited before anything is published. Live: [kernelbench.com](https://kernelbench.com).

This monorepo is the website **and** the evals. GPU sessions launch to Lambda / Brev (or another remote worker). Operator workflow: `AGENTS.md`. Methodology and history: each bench's `SPEC.md` and `DEVLOG.md`.

Benches

bench	path	what	site
hard	benchmarks/hard/	per-op kernels (CUDA or Triton), roofline-graded	/hard
mega	benchmarks/mega/	full fused megakernels	/mega
cuda	benchmarks/cuda/	CUDA-only writing deck (Triton/DSL fail)	/cuda
mini	benchmarks/mini/	small-model (<200B) deck, capped + 5-repeat (WIP)	unpublished
multi	benchmarks/multi/	4×H100 NVLink multi-GPU (WIP, frontier roster)	unpublished
v3	benchmarks/v3/	offline archive (separate harness)	not on site

Hard / mega / cuda share harness machinery and run unlimited wall-clock (`BUDGET_SECONDS=0`). Mini is capped; multi is sequential on a 4-GPU node.

Website (local)

Next.js 16 + Tailwind. Package manager is **bun** (`bun.lock`):

bun install
bun dev          # http://localhost:3000
bun run build

Site data is baked at build time from `benchmarks/*/results/` (`app/_lib/data.ts`). Publish/deploy: `kb publish` then `kb deploy` — see `AGENTS.md`.

Docs map

• AGENTS.md — sweeps, harnesses, Lambda/Brev, publish, audits
• benchmarks//SPEC.md — methodology
• benchmarks//DEVLOG.md — design history
• benchmarks//README.md — short human entry for that deck