Beyond Experience: Knowledge-Guided Evolution for NPU Kernel Generation

Key points

• LLM agents have become capable kernel engineers on mainstream GPU stacks, but they struggle on emerging accelerators such as Ascend NPUs, whose programming model is barely represented in pretraining data.1
• Recent NPU agents address this gap largely through experience: they accumulate memory from their own generation and optimization trials and reuse it on later tasks.1
• We argue that experience alone is not enough for kernel generation.1
• On 53 Ascend C operators from CANN-Bench spanning four difficulty levels, ASCENDEVOLVE achieves an overall score of 71.39, improving over Reflexion and OpenEvolve by 11.6% and 17.2%.1

Sources (1)

• [1]Beyond Experience: Knowledge-Guided Evolution for NPU Kernel Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 03:56 AM

  We present ASCENDEVOLVE, which grounds every step of an evolutionary kernel search in a large external knowledge base built from platform documentation, toolkit headers, and operator repositories.

  LLM agents have become capable kernel engineers on mainstream GPU stacks, but they struggle on emerging accelerators such as Ascend NPUs, whose programming model is barely represented in pretraining data.

Extractive summary: sentences quoted from the sources.