Task and objective
Improve a representative grouped matrix multiplication kernel while preserving numerical correctness across the published workload. This page demonstrates the specificity expected from a sponsor; it is not a live bounty or hardware result.
Primary metric & scoring overview
Illustrative primary metric: median isolated kernel time after the correctness gate. Integration acceptance remains separate and mandatory; no score or result is claimed here.
Workload envelope
Evaluate grouped matrix multiplications across published small, medium, and large shape sets using FP16 inputs and FP32 accumulation. A live contract would enumerate every shape, distribution, seed, and excluded case.
Environment
Illustrative only: NVIDIA H100 SXM, a pinned CUDA and compiler version, fixed clocks, and an identified host configuration. These are example terms, not a report of available hardware.
Correctness gate
Compare every published case with a named reference implementation using declared absolute and relative tolerances. Any timeout, crash, unsupported shape, or out-of-tolerance output fails the gate.
Measurement
After warm-up, synchronize around repeated runs and rank by the stated median kernel time. Report dispersion and rerun noisy cases. This describes a protocol; no measurements have been taken.
Integration acceptance
The source must compile in the published harness, preserve its callable interface, pass regression cases, and avoid worsening end-to-end latency. Isolated kernel speed does not by itself establish integrated value.
Resources & security
One UTF-8 source file, approved toolchain dependencies only, no network access, and explicit time and memory limits. A live sponsor must publish sanitized inputs and must not include secrets or raw telemetry.
Artifact scope
Submit one UTF-8 CUDA, C++, header, or Python source file with short build and design notes. Repository, archive, and container uploads are outside the current product scope.
Award & dispute
Among entries passing correctness and integration acceptance, the lowest primary metric wins. A live contract must define ties, invalidation, the sponsor review window, and dispute handling before posting.
Baseline
Reference link supplied when liveSubmitted entries
Example pages do not accept or display submissions.