KernelArc: A Multi-Agent Framework for GPU Kernel Optimization
06:00 · August 19, 2026 · arXiv cs.AI RSS
We present KernelArc, a multi-agent framework for autonomous GPU kernel optimization across heterogeneous workloads. Strategy-specialized agents run in parallel and coordinate through conclusions-only shared memory, a deterministic benchmark guard, and read-only cross-agent state with plateau-triggered drafting. We evaluate \kernelarc{} on NVIDIA H100 and B200 GPUs using category-representative SOL-ExecBench workloads. The resulting implementations span 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. At the public SOL-ExecBench leaderboard snapshot recorded on July~30, 2026, these submissions ranked first on representative L1, L2, Quantization, and FlashInfer tasks. The trajectories support the paper's central motivation: shared multi-agent search can broaden exploration and reach stronger incumbents within a fixed candidate budget, while the value of individual coordination features depends on the kernel and optimization stage.
Summary
KernelArc is a multi-agent framework that automates the search for high-performance GPU kernels on NVIDIA Hopper and Blackwell hardware. It addresses the difficulty of coordinating low-level features such as WGMMA instructions, TMA transfers, asynchronous barriers, and NVFP4 formats across memory layouts, register pressure, and synchronization constraints. Rather than relying on a single sequential agent, the system deploys multiple strategy-specialized agents that explore distinct optimization families in parallel.
Coordination occurs through a conclusions-only shared memory that retains only validated findings, a deterministic guard that enforces correctness checks and benchmark decisions, and read-only access to stronger solutions discovered by sibling agents. When an agent reaches a performance plateau, a trigger forces it to draft an alternative approach instead of continuing incremental refinement. This design keeps generative reasoning with the language models while confining state management, launch logic, and leaderboard tracking to deterministic code.
The framework was tested on representative workloads from the SOL-ExecBench suite, including custom BF16 GEMM kernels, static cuBLASLt configuration tables, fused mixture-of-experts backward passes, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention, and paged prefill attention. On the public leaderboard snapshot of 30 July 2026, the generated kernels ranked first in the L1, L2, Quantization, and FlashInfer categories. Ablation trajectories indicate that shared multi-agent search expands coverage within a fixed evaluation budget compared with single-agent baselines, although the benefit of each coordination mechanism varies with the kernel and the stage of optimization.
Why it matters
High technical depth and novelty in multi-agent kernel search; directly actionable for Dutch AI/HPC teams working on performance engineering; IMEC affiliation adds EU relevance for advanced GPU workloads.
