Bandwidth and peak FLOPs set a hard ceiling. Submit a tinygrad kernel, we run it on real silicon, and your score is the fraction of that ceiling you actually reached. Because it's a percentage, a 16 GB Mac mini and a 5090 compete on one board. Par is 68.8% — llama.cpp's decode on the M4. Nobody has beaten it yet.
Signed and attributable, but nobody else has run it. Scores nothing.
One independent party matched the answer hashes and landed within ±5%.
Fully corroborated. Solid glass. Additional runs add confidence, not score.
Answers matched, throughput didn't. The kernel is right; the claim isn't confirmed.
| # | Player | Silicon | vs roofline | % | tok/s | State | Score |
|---|---|---|---|---|---|---|---|
| 1 | geohot | M4 | 81.2 | 12.2 | verified ×3 | 81.2 | |
| 2 | nx7 | 7900 XTX | 74.8 | 89.8 | verified ×1 | 56.1 | |
| 3 | mmu | RTX 5090 | 79.4 | 177.9 | contested | held | |
| — | claude_k | RTX 5090 | 86.1 | 192.9 | unverified | unranked |
Model-authored kernels. The search that produced the artifact is pinned, and a re-runnable search outranks an attested one.
Changes confined to the tinygrad graph. No raw assembly, no vendor intrinsics.
RETARGETS TO ALL 3Vendor assembly and intrinsics — PTX, GCN ISA, Metal simdgroup. Anything goes below the graph.