KernelArc: A Multi-Agent Framework for GPU Kernel Optimization
- Source
- arxiv.org
- Author
- Joyjit Kundu, Ben Stoffelen, Kaili Wang, Peter Vrancx, Ludovic Denoyer
- Date

Context
KernelArc divides the search for faster GPU routines among agents pursuing different optimization strategies. They share conclusions, can inspect one another's work without modifying it, and use a common guard. This lets parallel searches exchange useful findings while preserving separate experiments. The paper evaluates the approach on NVIDIA H100 and B200 hardware.
The authors report first-place submissions on selected SOL-ExecBench tasks in the July 30, 2026 leaderboard snapshot. Their stronger general finding is conditional: collaboration broadened exploration within a fixed number of candidate implementations, but the value of individual coordination features depended on the routine and optimization stage.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
articleHPCAgent-Bench: A Framework for Sound Benchmarking of Code-Optimizing Agents, Harnesses, and SkillsYakup Koray Budanaz, Lorenzo Paleari, Marcin Copik, Alexandru Calotoiu, Timo Schneider, Benjamin Chen Ming Choong, Shriram Chandran, Lars Hulsbergen, Levi Widmer, Torsten Hoefler
articleDataKernelBench: Can LLMs Optimize Database Queries on GPUs?Gokul Karthik Kumar, Yotam Perlitz, Corey Lammie, Andrea Giovannini, Katja Hose
articleDo AI Agents Understand Computer Architecture?Ambika Sharan, Grigory Chirkov, Soheil Abbasloo
Checking sign-in…
Loading comments…