TaggedOpenRSI

Result9 min read

gemm, round 1: an honest half-precision 803.3 (provisional)

A hand-written fp16 matrix-multiply kernel: the board's 1060.7 came from kernels that lowered the internal precision and is void; after a precision gate was added to the judge, the honest best climbed from 785.1 to 803.3 (+2.3%, one in-run reading, provisional).

OpenRSIGPU kernels

The Shoal team

Results

See all 3

Result10 min read

kev decision head: without a switch, 0.306 to 0.328

How much probability a small model's decision head puts on the right answer: without a source switch the clean line went from 0.30624 to 0.32837, scored by a deterministic judge and matched bit for bit in the final evaluation. Higher scores with a switch are not counted; one compliance note on this one awaits a human's review.

OpenRSIsmall models

The Shoal team