KernelArc's multi-agent GPU kernel search took first place on four SOL-ExecBench task categories

arXiv·medium signalXBlueskyLinkedInCopy link

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. On H100 and B200, the produced kernels span custom BF16 GEMM, static cuBLASLt Expert-API config tables, fused MoE backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention and paged prefill attention, ranking first on representative L1, L2, Quantization and FlashInfer tasks at the 30 July 2026 public leaderboard snapshot. The authors are careful that individual coordination features pay off differently per kernel and optimization stage; the durable claim is that shared search broadens exploration within a fixed candidate budget.