AMD-AGI/AgentKernelArena
76 stars · Last commit 2026-07-10
AgentKernelArena provides an end-to-end siloed-benchmarking environment where different LLM-powered agents—such as Cursor Agent, Claude Code, Codex, SWE-agent, and GEAK—can be evaluated side-by-side on the same GPU kernel tasks, using objective and reproducible metrics.
README preview
# AgentKernelArena: Competitive Arena for GPU Kernel Optimization Agents As AI coding agents, like Claude Code and OpenAI Codex, rapidly improve, we need more than cherry-picked demos. Especially in specialized domains like GPU programming. AgentKernelArena is a standardized evaluation arena built by AMD to measure how well AI coding agents perform on real GPU kernel optimization tasks. ## Demo A live illustrative demo is available at: http://165.245.130.75/ This demo is provided only to illustrate the results and should not be treated as the final benchmark leaderboard. ## Overview & Features AgentKernelArena provides an end-to-end, siloed benchmarking environment where LLM-powered agents (Cursor Agent, Claude Code, Codex, and custom agents) are evaluated side-by-side on the same kernel tasks using objective and reproducible metrics. AgentKernelArena enables systematic evaluation of AI agents on GPU kernel optimization tasks by combining: - **Multi-Agent Arena**: Cursor, Claude Code, Codex, and custom agents - **Multi-Model Support**: OpenAI (GPT-5), Anthropic Claude (Opus and Sonnet families), and other models via OpenRouter or vLLM - **Task Categories**: HIP (ROCm examples, rocPRIM, customer HIP), Triton (vLLM-style local harnesses and ROCmBench), and Torch2HIP conversions