AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI Agents Search for Faster Blackwell GPU Kernels, Reaching the Top of Three Public Boards

Speed-critical GPU code is still written largely by hand, and new hardware features often outpace the libraries that use them. A new paper from Amazon researchers describes CompoundingKernel, an agent system that automates that work by searching. As of November 11, 2026, its generated kernels ranked first on three public boards of SOL-ExecBench, NVIDIA’s benchmark for kernels written by AI agents on NVIDIA’s B200 GPUs.

A kernel is a program that runs a single operation on a graphics processor. The authors’ search starts each problem from its PyTorch definition, with no reference kernels and no official scores. Each problem has an archive, a database outside the language model that records every correct kernel and its parent. An LLM controller, Claude Opus 5.5 in these runs, reads the archive, chooses a parent kernel or opens a new design, and gives each of several parallel child agents one direction to try. Each child edits, compiles and times its code. Checks outside the model decide correctness, and slower correct kernels stay in the archive, where they can become the starting point for later work.

The paper gives one example of how this can help. On a quantized decoder-layer problem, the best kernel’s local speedup rose from 6.25× to 27.36× over the official baseline. At one point, a child replaced an existing attention component with a version written for Blackwell’s tensor cores. That first attempt ran at 1.52×. A later child traced the slowdown to register spills and reached 7.10×, and a third child moved the new attention onto the best kernel at the time, reaching 13.97×. The authors report that 93% of the run’s nodes descend from the 1.52× version.

On the site’s timing, the kernels are 4.40× faster than the benchmark’s baseline across 209 public-board problems (geometric mean; median 3.78×). Board-level geometric means were 4.04× on L1, 3.62× on L2 and 9.19× on Quantization. The official mean scores were 0.823339 (L1), 0.819260 (L2) and 0.909670 (Quantization), with a score of 1.0 representing the speed-of-light bound. The team also reports a site-verified correct kernel for all 235 problems. In a controlled comparison, the LLM controller beat a coded tree search on 35 of 38 paired problems at equal wall-clock time, mainly by making 2.4× as many correct attempts. Per attempt, the authors found the two picked about equally well. A warm restart, which updated the harness while keeping the archives, raised evaluated candidates per hour 3.6×, although the restart bundled several changes. Making reference documents mandatory produced no measurable gain, with a geometric-mean ratio of 0.981 over 233 pairs.

The results have important caveats. Local timings ran on GPUs with unlocked clocks and read faster than the site’s; the authors correct for this when choosing what to submit. Two problems ended slower than baseline, and 21 of the 209 public problems were below 1.5×. The evidence comes from one benchmark, one model and one cold-started final run, and several design comparisons used earlier harness versions with a single run per arm. The checks were not applied to every candidate. Humans chose which kernels to submit and allocated compute using public standings, though agents never saw scores. The authors’ view that the rate of reliable measurements, rather than agent count, limits scaling is not yet tested directly.

The work suggests that automated kernel search can gain most from fast, trustworthy feedback on each attempt, and that a slower kernel can still be a useful stepping stone. It does not establish that the kernels are correct on all inputs, or that the approach transfers to other hardware or benchmarks. The authors plan to test cheaper verification and other GPUs.