KernelArc multi-agent framework tops GPU kernel benchmarks with parallel AI agents

KernelArc, introduced in an August 2026 arXiv paper by Joyjit Kundu, Ben Stoffelen, Kaili Wang, Peter Vrancx, and Ludovic Denoyer, is a multi-agent framework that automates GPU kernel optimization across heterogeneous workloads. Unlike single-agent tuners, KernelArc deploys strategy-specialized agents that explore different optimization paths in parallel. They coordinate through a conclusions-only shared memory, a deterministic benchmark guard to ensure correctness, and read-only cross-agent state with plateau-triggered drafting—allowing agents to build on each other's findings without costly full-state exchanges. This design broadens exploration within a fixed candidate budget and helps reach stronger incumbents than traditional search.

Evaluated on NVIDIA H100 and B200 GPUs using the SOL-ExecBench suite, KernelArc generated production-quality implementations including custom BF16 GEMM, static cuBLASLt Expert-API configuration tables, fused mixture-of-experts backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention, and paged prefill attention. On the July 30, 2026 SOL-ExecBench leaderboard snapshot, its submissions ranked first across representative L1, L2, Quantization, and FlashInfer tasks. The paper highlights that shared multi-agent search enables broader exploration and stronger results, though the value of individual coordination features depends on the kernel and optimization stage—a nuanced finding for future research in automated performance engineering.