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

This paper introduces ChatNeuroSim, an LLM-based agent framework that automates the entire Compute-in-Memory (CIM) accelerator deployment and optimization workflow—including simulation, parameter management, and design space pruning—to significantly reduce the time and effort required for design space exploration.

The Gist

Imagine you are trying to build the most efficient, fastest, and cheapest smart kitchen possible. But instead of a normal kitchen, this one uses a revolutionary new way of cooking called "Compute-in-Memory" (CIM).

In a normal kitchen (traditional computers), the chef (processor) has to run back and forth to the pantry (memory) to grab ingredients, cook them, and run back. This running back and forth wastes time and energy. In a CIM kitchen, the pantry is the stove. The ingredients are cooked right where they are stored. It's incredibly fast and efficient, but designing one is a nightmare.

Here is the story of ChatNeuroSim, the new tool that makes designing these super-kitchens easy.

The Problem: The Confusing Instruction Manual

Designing a CIM kitchen used to be like trying to build a spaceship using a 500-page manual written in a foreign language.

• The Manual: You had to read thousands of pages of technical specs to understand how the stove, the pantry, and the chef interact.
• The Trial and Error: You'd guess a setting, run a simulation (a test cook), see if it burned the food, and then guess again. This took days or weeks.
• The Result: Even experts got stuck, and it was too slow to build the next generation of AI hardware.

The Solution: ChatNeuroSim (Your AI Sous-Chef)

The researchers built ChatNeuroSim, which acts like a super-smart AI Sous-Chef that speaks your language. You don't need to read the manual anymore; you just talk to the AI.

How it works (The Three Agents):
Think of ChatNeuroSim as a team of three specialized robots working together:

1. The Translator (Task Parsing Agent):

   • You say: "I want to cook a Swin Transformer dish with low power usage."
   • The Robot: "Got it! You want a 'PPA Optimization' request. I know exactly what that means."
   • It translates your vague human wish into a specific category of task.

2. The Inventory Manager (Parameter Parsing Agent):

   • You say: "Use 22nm technology and 8-bit precision."
   • The Robot: "Okay, but you didn't tell me the size of the pantry or the type of stove. Let me check the master list (the database) and ask you for those missing details."
   • It fills in the blanks, checks if your numbers make sense, and organizes the ingredients perfectly.

3. The Head Chef (Execution Agent):

   • Once everything is ready, this robot runs the actual simulation. It tells the computer to "cook" the design, checks the results, and tells you: "Here is your recipe, and here is how much energy it saves."

The Secret Sauce: The "Smart Filter" (Design Space Pruning)

Even with a great AI, searching for the perfect kitchen design is like looking for a needle in a haystack the size of a city. There are billions of possible combinations of stove types, pantry sizes, and chef speeds.

The paper introduces a clever trick called Design Space Pruning.

• The Analogy: Imagine you want to design a kitchen for a new type of cuisine (a Vision Transformer). Instead of starting from scratch and testing every single stove in the world, the AI looks at a kitchen it already built for a similar cuisine (a ResNet).
• The Magic: It says, "Hey, we know that for the similar cuisine, big stoves with small pans worked best. Let's ignore all the tiny stoves and giant pans for this new job. Let's only test the 'big stove' options."
• The Result: It cuts away 60% to 80% of the useless options before it even starts cooking. This makes the search 2x to 3x faster.

The Results: What Did They Achieve?

The team tested this system with 40 different requests (from simple questions to complex optimization tasks).

• 100% Success Rate: When using the best AI model, the system got the instructions right every single time. No more "I misunderstood your manual" errors.
• Speed Boost: For the hardest designs (like the Vision Transformer), the "Smart Filter" cut the time needed to find the best design from hours down to minutes.
• Better Designs: Not only was it faster, but the designs it found were often better (more energy-efficient) than what humans found by guessing.

Why Should You Care?

We are moving toward a future where AI runs everything—from self-driving cars to your smart fridge. These AI systems need hardware that is fast and doesn't overheat.

ChatNeuroSim is the tool that stops engineers from drowning in paperwork and manual testing. It turns a process that used to take weeks of human effort into a chat conversation. It allows us to build the "super-kitchens" of the future much faster, cheaper, and smarter.

In short: It's the difference between trying to build a car by reading a dictionary of engine parts versus just telling a robot, "Build me a fast, fuel-efficient car," and watching it do the rest.

Technical Summary

Here is a detailed technical summary of the paper "ChatNeuroSim: An LLM Agent Framework for Automated Compute-in-Memory Accelerator Deployment and Optimization."

1. Problem Statement

Compute-in-Memory (CIM) architectures are critical for accelerating Deep Neural Networks (DNNs) by reducing data movement overhead. However, the design and optimization of CIM accelerators face two major bottlenecks:

• Usability and Complexity: System-level simulators like NeuroSim require designers to manually interpret complex documentation, understand intricate parameter dependencies, and convert high-level design intents into specific simulator inputs. This iterative process is time-consuming and prone to human error.
• Inefficient Design Space Exploration (DSE): The CIM design space is vast and highly coupled (often $10^3$–$10^4$ dimensions). Identifying optimal hardware configurations under strict Power, Performance, and Area (PPA) constraints requires extensive manual tuning or repeated simulations. Traditional heuristic optimization algorithms (e.g., Genetic Algorithms, Simulated Annealing) often suffer from high runtime costs, particularly for complex workloads like Vision Transformers (ViTs), which have larger design spaces than Convolutional Neural Networks (CNNs).

2. Methodology

The authors propose ChatNeuroSim, a Large Language Model (LLM)-based multi-agent framework that automates the entire CIM workflow, coupled with a novel Design Space Pruning technique for optimization.

A. The ChatNeuroSim Agent Framework

The framework consists of three coordinated LLM agents that interact with the NeuroSim simulator:

1. Task Parsing Agent: Classifies user requests into four categories:
   • Single Call: One simulation with fixed parameters.
   • Multiple Call: Multiple simulations with explicit parameter sets.
   • Testbench Auto-design: Multiple simulations where the agent determines parameter sweeping values based on high-level intent.
   • PPA Optimization: Finding optimal configurations under constraints.
2. Parameter Parsing Agent: Extracts structured parameters from natural language queries using Retrieval-Augmented Generation (RAG) to consult simulator schemas. It validates parameters against supported values, units, and dependencies.
3. Parameter Adjustment Agent: Handles user feedback to modify, add, or remove parameters, ensuring consistency across specialized and common parameters before script generation.

The system uses Prompt Engineering to enforce strict rules, prevent hallucinations, and ensure deterministic tool invocation. It generates executable scripts, runs NeuroSim, and returns PPA metrics.

B. CIM Optimizer with Design Space Pruning

To address the high cost of DSE, the authors introduce a heuristic ML-based optimizer enhanced by Transfer Learning and Design Space Pruning:

• Cross-Space Constraint Projection: Uses power-law scaling ($Y = \alpha X^\beta$) to map hardware constraints (e.g., power, area) from a "base model" (e.g., ResNet-50 or ViT-B) to a "target model" (e.g., Swin-T). This allows the system to predict feasibility without running expensive simulations for every candidate.
• Top-K Pruning: Before optimization begins, the design space is pruned by retaining only parameter values (binned into small, mid, large) that correspond to the Top-K performing configurations of the base model.
• Stochastic De-pruning: To avoid getting trapped in local optima, the system periodically restores excluded parameter values stochastically during the optimization process, balancing exploitation of the pruned space with exploration of the full space.

3. Key Contributions

1. Automated CIM Workflow: A fully automated LLM agent framework that handles request parsing, parameter validation, script generation, and simulation execution, eliminating the need for manual simulator configuration.
2. Efficient Optimization Strategy: A novel CIM optimizer that integrates design space pruning with transfer learning. It leverages knowledge from prior workloads to accelerate the search for optimal configurations in new, complex workloads.
3. Comprehensive Evaluation:
   • Validated the agent framework on a custom dataset of 40 CIM requests, achieving 100% correctness in script generation with GPT-5.1.
   • Demonstrated significant runtime reductions in optimization tasks compared to baseline methods without pruning.

4. Experimental Results

The framework was evaluated using NeuroSim V1.5 on a 22nm technology node with various DNN workloads (ResNet-50, Swin-T, ViT-B).

• Agent Accuracy: Using GPT-5.1, ChatNeuroSim achieved 100% accuracy across all 40 test cases (Single Call, Multiple Call, Auto-design, and PPA Optimization). Other models like GPT-5-mini and DeepSeek showed lower accuracy (77.5%–95%), highlighting the need for robust prompting and model selection.
• Optimization Runtime Reduction:
  • For Swin Transformer Tiny (Swin-T) optimization, the proposed design space pruning reduced the average runtime by 0.42x–0.79x and the P95 runtime by 0.29x–0.69x compared to baseline algorithms without pruning.
  • Simulated Annealing (SA) consistently outperformed Genetic Algorithms (GA) and Tree-Structured Parzen Estimators (TPE) in convergence speed.
• Base Model Selection: Pruning based on ViT-B (a similar architecture to Swin-T) yielded consistent speedups across all constraints and objectives. Pruning based on ResNet-50 was less effective for ViT workloads due to architectural differences.
• Solution Quality: The pruning strategy not only accelerated the process but also improved the quality of solutions (higher FoM) when optimization iterations were limited.

5. Significance

• Democratization of CIM Design: ChatNeuroSim lowers the barrier to entry for CIM research by allowing designers to interact with complex simulators using natural language, removing the steep learning curve associated with simulator manuals and parameter dependencies.
• Accelerated DSE: The design space pruning technique provides a systematic method to reduce the computational cost of exploring complex hardware configurations, which is critical for the rapid deployment of next-generation AI accelerators.
• Scalability: The framework is designed to handle diverse workloads, from CNNs to Vision Transformers, making it a versatile tool for the evolving landscape of AI hardware.
• Open Source: The authors have open-sourced the framework and a user interface, enabling the research community to trial and extend the system.

In conclusion, ChatNeuroSim represents a significant step forward in EDA automation, successfully integrating LLM agents for usability and advanced ML techniques for optimization efficiency in the domain of Compute-in-Memory architectures.