ChatNeuroSim: An LLM Agent Framework for Automated Compute-in-Memory Accelerator Deployment and Optimization

2026-03-11 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Hardware Design & Optimization, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

ChatNeuroSim is a large language model (LLM)-based agent framework designed to automate the deployment and optimization of Compute-in-Memory (CIM) accelerators. It addresses challenges in conventional CIM design flows, such as complex simulator manuals and extensive design-simulation iterations, by automating tasks like request parsing, parameter dependency checking, script generation, and simulation execution. The framework integrates a CIM optimizer that utilizes design space pruning, which significantly accelerates the identification of optimal CIM configurations for various deep neural network (DNN) workloads. Evaluated on 40 request-level testbenches, ChatNeuroSim achieved 100% accuracy in script generation and simulation results using GPT-5.1. A case study optimizing Swin Transformer Tiny under 22 nm technology demonstrated that the proposed CIM optimizer with design space pruning reduced average runtime by 0.42x–0.79x compared to a no-pruning baseline.

Key takeaway

For AI Scientists and Research Scientists engaged in Compute-in-Memory (CIM) accelerator design, ChatNeuroSim offers a significant reduction in design cycle time and manual effort. You should consider integrating this LLM-based framework to automate complex design space exploration, especially for vision transformer workloads like Swin-T, where it has shown substantial runtime reductions. Leveraging its design space pruning capabilities can accelerate identifying optimal hardware configurations, allowing you to focus on higher-level architectural innovations.

Key insights

ChatNeuroSim automates CIM accelerator design and optimization using LLM agents and design space pruning for faster, more efficient exploration.

Principles

• LLM agents can automate complex EDA workflows.
• Design space pruning accelerates hardware optimization.
• Transfer learning improves search efficiency across models.

Method

ChatNeuroSim employs three LLM agents (task parsing, parameter parsing, adjustment) and a CIM optimizer with cross-space constraint projection, Top-K pruning, and stochastic de-pruning to automate DSE.

In practice

• Use ChatNeuroSim for automated CIM accelerator design.
• Apply design space pruning for vision transformer optimization.
• Select base models with similar architectures for pruning.

Topics

• Compute-in-Memory
• LLM Agents
• Design Space Exploration
• Hardware Optimization
• Vision Transformers

Code references

• neurosim/ChatNeuroSim
• langchain-ai/langgraph
• streamlit/streamlit

Best for: AI Scientist, Research Scientist, AI Engineer, AI Architect, AI Researcher

Related on AIssential

• Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization — LLM agents in code optimization primarily leverage pretrained priors over iterative feedback or agentic exploration.
• ChatSVA: Bridging SVA Generation for Hardware Verification via Task-Specific LLMs — ChatSVA significantly improves SVA generation by decomposing reasoning and synthesizing high-purity, domain-specific data for LLMs.
• Agentic Architect: An Agentic AI Framework for Architecture Design Exploration and Optimization — LLM-driven code evolution combined with cycle-accurate simulation can discover novel, high-performing microarchitectural designs.
• Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems — Latent communication in LLM multi-agent systems reduces cost and information loss by exchanging continuous representations directly.
• There Is No AI. There’s a Stateless Function on 10,000 GPUs Pretending to Know You (Ep. 299) — LLM "magic" is sophisticated engineering, not inherent model intelligence, relying on stateless design and clever caching.
• Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention — LLM architectures are evolving with complex, targeted tweaks to optimize long-context efficiency and reduce computational overhead.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.