SimulatorCoder: DNN Accelerator Simulator Code Generation and Optimization via Large Language Models

TL;DR

SimulatorCoder leverages large language models with advanced prompt engineering to automatically generate and optimize DNN accelerator simulators, achieving high accuracy and efficiency with minimal manual effort.

Contribution

This work introduces a novel LLM-based framework for automated DNN simulator code generation and optimization, integrating multi-faceted prompt strategies and feedback loops.

Findings

Achieves less than 1% error in cycle-level fidelity compared to manual simulators.

Significantly reduces simulation runtimes while maintaining accuracy.

Structured prompting improves code correctness and simulator performance.

Abstract

This paper presents SimulatorCoder, an agent powered by large language models (LLMs), designed to generate and optimize deep neural network (DNN) accelerator simulators based on natural language descriptions. By integrating domain-specific prompt engineering including In-Context Learning (ICL), Chain-of-Thought (CoT) reasoning, and a multi-round feedback-verification flow, SimulatorCoder systematically transforms high-level functional requirements into efficient, executable, and architecture-aligned simulator code. Experiments based on the customized SCALE-Sim benchmark demonstrate that structured prompting and feedback mechanisms substantially improve both code generation accuracy and simulator performance. The resulting simulators not only maintain cycle-level fidelity with less than 1% error compared to manually implemented counterparts, but also consistently achieve lower simulation runtimes, highlighting the effectiveness of LLM-based methods in accelerating simulator development. Our code is available at https://github.com/xiayuhuan/SimulatorCoder.