Back to articles

CompoundingKernel Turns GPU Kernel Tuning into a Search Loop

2 min read

Generating a GPU kernel is only the first step: a useful system must also establish correctness and keep searching for faster implementations. CompoundingKernel, presented in the paper, treats kernel writing as an iterative search. Each successful implementation is retained, so later attempts can build on previous work rather than start from scratch.

For each problem, agents receive a PyTorch definition, but not reference kernels or official benchmark scores. An archive kept outside the model stores every correct kernel and its lineage. A language-model controller consults this record, chooses an existing design as a parent or starts from a new design, and gives parallel child agents distinct directions to explore. Correctness is decided by checks outside the model, rather than by the code generator itself.

The paper reports that, as of November 11, 2026, CompoundingKernel ranked first on the public L1, L2, and Quantization boards of SOL-ExecBench on NVIDIA B200. All 235 problems had a kernel verified as correct on the site. On the 209 problems represented on public boards, the kernels were 4.40× faster than the benchmark reference baseline. The authors also report that the system held first place on L1 from October 27, with a brief loss of the lead on November 7.

Evaluation throughput, not just more agents

In an equal-wall-time comparison with a coded UCB-style search tree, CompoundingKernel produced more correct attempts on 35 of 38 problems, reaching 2.4 times as many. The system ran more than a thousand agents concurrently. Separately, a fix to warm restarts increased its evaluation rate by 3.6 times. The authors’ central scaling observation is that the supply of reliable measurements limits how far search can grow; adding agents alone does not remove that bottleneck.

The broader idea is to make kernel optimization cumulative. A lineage-tracked archive can preserve useful discoveries, while external checks provide a clear filter for candidate implementations. The results also highlight evaluation infrastructure as a core part of agentic coding systems, not merely a final scoring step. Still, the reported standings and speedup are tied to a particular benchmark, accelerator, and snapshot in time. They do not establish that the same gains will carry over to every workload or hardware platform. Results across other tasks and independent reproductions will help clarify how general the approach is.

Source: Hugging Face Daily Papers

Comments

Checking sign-in status...

Loading comments...

Related articles