AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents

2026-05-16

AI Summary

Existing GPU kernel optimization benchmarks evaluate only single large-model invocations and lack assessment of full-agent workflows and cross-configuration generalization. To address this gap, this work proposes AgentKernelArena—an open-source benchmark that introduces, for the first time, a generalization evaluation protocol for unseen input configurations, enabling end-to-end assessment of AI programming agents across multiple backends (e.g., HIP, Triton, PyTorch) and tasks (optimization and translation). The benchmark integrates isolated workspaces, gated compilation, performance profiling, and centralized scoring. Experiments show that leading agents achieve an average speedup of 6.89×; generalization tests further reveal significant shape dependency in PyTorch-to-HIP translation tasks, while HIP and Triton kernel optimizations demonstrate strong generalization capabilities.

Abstract

GPU kernel optimization is increasingly critical for efficient deep learning systems, but writing high-performance kernels still requires substantial low-level expertise. Recent AI coding agents can iteratively read code, invoke compilers and profilers, and refine implementations, yet existing kernel benchmarks evaluate single LLM calls rather than full agent workflows, and none include both kernel-to-kernel optimization and unseen-configuration generalization testing. We present AgentKernelArena, an open-source benchmark for measuring AI coding agents on GPU kernel optimization. The benchmark contains 196 tasks spanning HIP-to-HIP optimization, Triton-to-Triton optimization, and PyTorch-to-HIP translation, and evaluates complete agent workflows in isolated workspaces using gated compilation, correctness, and performance checks, centralized scoring and an unseen-configuration generalization protocol that tests whether optimizations transfer to input configurations the agent never observed. Across production agents including Cursor Agent, Claude Code, and Codex Agent, we find near-perfect compilation and high correctness rates on most task categories, with the strongest configurations achieving mean speedups of up to 6.89x on PyTorch-to-HIP, 6.69x on HIP-to-HIP, and 2.13x on Triton-to-Triton tasks. Our unseen-configuration evaluation shows that HIP-to-HIP and Triton-to-Triton optimizations largely transfer to unseen input shapes, while PyTorch-to-HIP exhibits substantial correctness drops, indicating that agents generating kernels from scratch frequently hardcode shape-specific assumptions. AgentKernelArena is designed as a modular, extensible framework for rigorous evaluation of agentic GPU kernel optimization across agents, tasks, and hardware targets.

Problem

• GPU kernel optimization
• AI coding agents
• generalization
• benchmarking
• unseen-configuration

Innovation

• GPU kernel optimization
• AI coding agents
• generalization evaluation
• agent benchmarking
• unseen-configuration transfer