---
title: "Speeding up GPU kernels by 38% with a multi-agent system"
description: "The post reports a three-week experiment in which a multi-agent system optimized 235 CUDA kernel problems generated from workloads spanning more than 124 production open-source models. Using SOL-ExecBench on 27 NVIDIA Blackwell 200 GPUs, the system used a planner to distribute and rebalance work, while autonomous workers benchmarked, debugged, and refined kernels without developer intervention. It ran separate CUDA C with inline PTX and CuTe DSL runs, reaching assembly-level optimization and learning a documented higher-level API. Against a single-agent optimized PyTorch baseline, it outperformed 149 of 235 problems, achieving a 1.38x geometric mean ratio, while 45 solutions exceeded 2x. The highlighted cases included an 84% geomean gain over a human FlashInfer baseline for paged-prefill attention, a 39% gain for NVFP4 MoE linear, and a GEMM reaching 86% of a tuned cuBLAS baseline; median SOL was 0.56, leaving room for improvement."
---

# Speeding up GPU kernels by 38% with a multi-agent system

[Cursor](https://yomu.fyi/company/cursor) · Wilson Lin, Sahil Modi, Yuan Zhang & Edward Lin · Apr 14, 2026

**Type:** Benchmark

## Summary

The post reports a three-week experiment in which a multi-agent system optimized 235 CUDA kernel problems generated from workloads spanning more than 124 production open-source models. Using SOL-ExecBench on 27 NVIDIA Blackwell 200 GPUs, the system used a planner to distribute and rebalance work, while autonomous workers benchmarked, debugged, and refined kernels without developer intervention. It ran separate CUDA C with inline PTX and CuTe DSL runs, reaching assembly-level optimization and learning a documented higher-level API. Against a single-agent optimized PyTorch baseline, it outperformed 149 of 235 problems, achieving a 1.38x geometric mean ratio, while 45 solutions exceeded 2x. The highlighted cases included an 84% geomean gain over a human FlashInfer baseline for paged-prefill attention, a 39% gain for NVFP4 MoE linear, and a GEMM reaching 86% of a tuned cuBLAS baseline; median SOL was 0.56, leaving room for improvement.

## Context

The experiment evaluated whether a long-running, multi-agent system could solve open-ended GPU kernel optimization problems beyond the manual, operation-by-operation tuning typically used by engineers. Kernel optimization provided measurable performance objectives without requiring a known target solution.

## Approach / What changed

NVIDIA used SOL-ExecBench to generate and benchmark 235 real-world problems from more than 124 production open-source models. A planner agent distributed and rebalanced work among autonomous workers, which learned to invoke the benchmarking pipeline for iterative testing, debugging, and optimization. Separate runs used CUDA C with inline PTX and CuTe DSL.

## Takeaways

- SOL-ExecBench evaluated solutions against software baselines and theoretical hardware limits on 27 NVIDIA Blackwell 200 GPUs, and invalidated results using tactics such as caching that exceeded what a B200 could support.
- For BF16 grouped-query attention with paged prefill, the system used hardware-specific memory and math instructions, persistent kernels, and input-size hyper-optimization to reach a 0.9722 SOL score and an 84% geomean speedup over FlashInfer.
- The system’s median SOL score was 0.56 despite the overall 38% geomean speedup, and the post says additional GPU compute could enable deeper exploration and further optimization.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [CUDA](https://yomu.fyi/topic/cuda), [GPU](https://yomu.fyi/topic/gpu), [Performance](https://yomu.fyi/topic/performance)

- Source: [Cursor](https://cursor.com/blog/multi-agent-kernels)
- Source URL: https://cursor.com/blog/multi-agent-kernels
- Ingested by Yomu: 2026-09-02T11:42:48.561Z

[Read original post](https://cursor.com/blog/multi-agent-kernels)