KernelArc: A Multi-Agent Framework for GPU Kernel Optimization

Details

GPU kernel optimization is traditionally a high-expertise manual process. KernelArc automates this by running strategy-specialized agents in parallel, coordinating through conclusions-only shared memory and plateau-triggered drafting. Cross-agent state is read-only, and a deterministic benchmark guard prevents spurious performance reports. The framework was evaluated on NVIDIA H100 and B200 GPUs using SOL-ExecBench workloads covering custom BF16 GEMM, cuBLASLt Expert-API configuration tables, fused mixture-of-experts backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention, and paged prefill attention. At the leaderboard snapshot on July 30, 2026, KernelArc submissions ranked first on L1, L2, Quantization, and FlashInfer tasks. The paper argues that shared multi-agent search broadens exploration beyond what a single agent can attempt within a fixed candidate budget. The value of individual coordination features — shared memory, plateau-triggered drafting — depends on the specific kernel and optimization stage rather than being uniformly beneficial. For ML infrastructure teams, KernelArc offers a concrete agentic approach to kernel tuning that complements or potentially replaces hand-optimization for common workloads on NVIDIA hardware.