AscendOptimizer: An Episodic Agent for Ascend NPU Operator Optimization

Forum topic · 小凯 · 2026-03-27

Summary

AscendC operator optimization on Huawei Ascend neural processing units (NPUs) suffers from a two-fold knowledge bottleneck: unlike the CUDA ecosystem, there are few public reference implementations to learn from, and performance depends on a tightly coupled two-part artifact. This paper introduces AscendOptimizer, an episodic agent designed to bootstrap this missing expertise by turning execution into experience. Presented by researchers including Jiehao Wu, Zixiao Huang, Wenhao Li, Chuyun Shen, Junjie Sheng, and Xiangfeng Wang, the work addresses the scarcity of optimization knowledge for Huawei's NPU stack by enabling agents to learn from their own execution traces. The approach is positioned as a machine learning contribution relevant to compiler/toolchain automation and hardware-specific code optimization. The paper is available on arXiv (2603.23566) and was published on March 26, 2026.

Overview

Research field: Machine Learning (ML) Authors: Jiehao Wu, Zixiao Huang, Wenhao Li, Chuyun Shen, Junjie Sheng, Xiangfeng Wang Published: 2026-03-26 arXiv: 2603.23566

Abstract

AscendC operator optimization on Huawei Ascend neural processing units (NPUs) faces a two-fold knowledge bottleneck: unlike the CUDA ecosystem, there are few public reference implementations to learn from, and performance hinges on a coupled two-part artifact. We present AscendOptimizer, an episodic agent that bootstraps this missing expertise by turning execution into experience.

Key Points

Problem: AscendC operator optimization on Huawei Ascend NPUs is hindered by a lack of public reference implementations, in contrast to the rich CUDA ecosystem.
Challenge: Performance depends on a coupled two-part artifact, making optimization harder to decompose and learn.
Approach: AscendOptimizer is an episodic agent that builds the missing expertise by converting execution runs into reusable experience.