Benchmark · Hard
The per-op kernel deck
Six problems × one unlimited agent session. Ranked by how many problems passed; each problem shows peak fraction or why it failed — fails are not averaged into a fake speed score.
methodology + notes
Reading the board. Order is pass count only. Green chips are peak fraction of that GPU's roofline. check = solution failed correctness; empty = no kernel written; hack = audit flag. "when ok" is mean peak fraction among passes only (footnote, not a rank key). Click a model for audits and traces. Run index.
RTX PRO 6000
%peak fractioncheckwrong / no passemptyno solutionhackaudit flag