Gemini 3.1 Pro

Google

One row per published session deck. Score bars are peak fraction of the board roofline (Hard / CUDA) or best speedup vs the torch baseline (Mega). Audit chips come from the human/subagent reward-hack review of every published cell.

hard

0/5

RTX PRO 6000

FP8 GEMMno run
KDA CUTLASSno run
Paged Attentionno run
TopK Bitonicno run
Sonic MoEno run
W4A16 GEMMno run

RTX 3090

FP8 GEMMno run
KDA CUTLASSfail
unaudited
session 47m
trace
Paged Attentionpass
61.2%unaudited
session 45m
trace
TopK Bitonicfail
unaudited
session 2h 7m
trace
Sonic MoEpass
15.5%unaudited
session 33m
trace
W4A16 GEMMpass
12.3%unaudited
session 34m
trace

Methodology: Rank per bench: valid passes (audited-clean correct cells / problems) desc, then mean normalized performance (cell score / board best per problem) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.