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

Researchers Ming-Yen Lee and Shimeng Yu have introduced ChatNeuroSim, a novel framework that uses large language model (LLM) agents to automate the complex process of designing and optimizing Compute-in-Memory (CIM) accelerators. CIM is a cutting-edge hardware architecture that boosts AI performance by performing calculations directly within memory, eliminating the data transfer bottleneck between memory and processors. Traditionally, engineers use simulators like NeuroSim for design space exploration (DSE), a tedious cycle of manual parameter tweaking and simulation that can take weeks. ChatNeuroSim tackles this by creating an AI agent that understands hardware specifications and neural network workloads, automatically generating and running the correct simulation scripts.

The framework's core innovation is its automation of the entire CIM workflow: parsing design requests, checking parameter dependencies, generating simulation code, and executing tasks. It also integrates a smart 'design space pruning' optimizer that narrows down the vast array of possible hardware configurations to find the optimal setup much faster. In a case study optimizing the Swin Transformer Tiny model for a 22nm technology node, ChatNeuroSim's optimizer reduced the average runtime by 42% to 79% compared to running the same optimization algorithm without its pruning technique. This validation demonstrates its effectiveness in parsing complex requests and executing correct simulations, moving AI hardware design from a manual, expert-driven process toward an automated, accelerated one.