AgentKernelArena - Triton-to-Triton: leaderboard

Metric: Mean speedup (x; arithmetic mean over the 148 Triton-to-Triton tasks (118 vLLM, 30 ROCmBench kernels) of the per-task speedup over the unmodified Triton kernel, counting 0 for a task whose kernel fails compilation or correctness; per-task mean of 3 runs; agent in an isolated workspace, 3600 s timeout, up to 3 kernel versions; AMD Instinct MI300X). Source: arxiv.org. Saturation forecast: Rough model projection: around 2027. 8 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=agentkernelarena-triton-to-triton · How It Works · Data refreshed daily, snapshot 2026-09-26.