KernelArc: Multi-Agent AI Framework Achieves Top GPU Kernel Optimization Results

The paper presents KernelArc, a framework designed for multi-agent autonomous optimization of GPU kernels across various workloads. In contrast to traditional single-agent methods, KernelArc utilizes specialized agents that work concurrently and communicate solely through shared memory conclusions. The framework features a deterministic benchmark guard and read-only state shared among agents, incorporating plateau-triggered drafting to improve exploration. Testing was performed on NVIDIA H100 and B200 GPUs utilizing representative SOL-ExecBench workloads. KernelArc's optimized outputs encompass a wide array of GPU tasks, such as custom BF16 GEMM, static cuBLASLt Expert-API configuration tables, and more. On July 30, 2026, the public SOL-ExecBench leaderboard showed these submissions leading in L1, L2, Quantization, and FlashInfer tasks, highlighting the effectiveness of shared multi-agent search in enhancing optimization results. This innovation marks a substantial leap in automated GPU kernel tuning, influencing high-performance computing and AI inference efficiency.