I run kernelbench.com, a live benchmark where frontier coding agents (Claude, GPT, Gemini, Kimi, DeepSeek, GLM, whatever ships next) write and optimize real GPU kernels. The scoring is against each GPU's hardware roofline instead of a soft PyTorch baseline, so a slow reference can't make anyone look good. There are three tracks right now: Hard is six per-op CUDA/Triton problems run across RTX PRO 6000, H100, and B200. Mega is fully fused megakernels, where Claude Fable 5 recently landed the first genuine single-launch megakernel at 18.7x over reference (Jack Clark covered it in Import AI 464). Multi is coming: NVLink collective kernels on 8xH100 nodes.

The part I care most about is that results are auditable. Every run publishes its full agent transcript to Hugging Face, and nothing goes on the board until it passes a judge-assisted reward-hacking audit. I have disqualified several impressive-looking "wins" this way. I also track output tokens per result, because a 15x kernel that cost 800k tokens is a different claim than one that cost 100k.

Why this matters: kernel engineering is one of the actual bottlenecks in AI research, and whether models can automate it is a question worth measuring carefully, independent of any lab grading its own homework.

What are this project's goals? How will you achieve them?

Over the 6 months: run full sweeps within days of every major model release (including API-priced models I currently skip on cost, like GPT-5.5 Pro), launch the Multi track publicly on rented 8xH100 nodes, and keep publishing transcripts plus per-run audits for everything on the board.

There is nothing speculative to build here. The harness, roofline machinery, audit pipeline, and transcript publication all exist and run today. Until now it has run on my coding-plan subscriptions and my own hardware, and that stopped scaling: Fable-class models at high effort are a different cost tier, some models are API-only, pricing moves around every release, and I don't own multi-GPU nodes. This funding takes a working benchmark from cost-limited to cost-covered.

How will this funding be used?

$15,000 over 6 months. Roughly: $6,000 on model credits/API so every major release gets a full sweep, $4,000 on multi-GPU cloud rentals for the Multi track (a full sweep at the 3h/run cap across the model matrix is somewhere around 100-150 node-hours on 8xH100), and $5,000 on my time, most of which actually goes into the per-run reward-hacking audits rather than running sweeps.

That split is indicative. Model pricing is volatile enough that the allocation will shift with real costs, and I will flag any substantial deviation to the funder.

Who is on your team? What's your track record on similar projects?

Just me. I built and run everything end to end: harness, problem decks, roofline scoring, audit pipeline, and the site. My day job is inference/training optimization at an AI avatar company; kernelbench happens entirely outside work hours.

Track record: Import AI 464 covered the benchmark. VentureBeat cited it as an independent authority when my run logs contradicted a lab's benchmark claims. The Mega track hosted the first verified single-launch megakernel result. Before this I did CUDA/Triton kernel work across NVIDIA, AMD MI300X, and TPU, RL research on training models to write faster kernels, and AI education with 2M+ learners through my FreeCodeCamp LLM and CUDA courses. Everything is public at github.com/Infatoshi.

What are the most likely causes and outcomes if this project fails?

The realistic failure mode is cost, not tech. If frontier pricing spikes past the credits budget, or 8xH100 availability makes Multi sweeps pricier than I estimated, the result is reduced cadence (fewer models or releases covered) or a delayed Multi launch. The site and existing boards stay up either way. The secondary risk is my time, since this lives outside a full-time job, which is exactly why the time line item exists.

How much money have you raised in the last 12 months, and from where?

$2,000 in hardware credits from Nvidia. Everything else has been self-funded through coding-plan subscriptions and my own hardware.

PS. A note on process, since a detector flagged an earlier version of this page: I use AI tools heavily in everything I do, including drafting text like this. The claims, numbers, and budget commitments are mine and I stand behind all of them. The benchmark itself is the evidence that matters - every run, transcript, and audit is public.