TLDR: AC-Refiner is a new AI framework that uses conditional diffusion models to efficiently optimize arithmetic circuits (like adders and multipliers) for better performance and smaller size. By treating circuit design as an image generation task and guiding it with desired performance metrics, it consistently produces superior designs, outperforming traditional and existing AI methods.
In the world of digital systems, arithmetic circuits like adders and multipliers are fundamental building blocks. Their efficiency directly impacts the performance, power consumption, and physical size of hardware. However, optimizing these circuits is a significant challenge due to the vast number of possible designs and complex physical limitations.
Traditional methods, including manual design and algorithmic approaches, often fall short in achieving optimal real-world performance because they struggle to accurately model physical constraints early in the design process. More recently, deep learning-based methods have emerged, but they too face limitations, particularly in efficiently exploring the massive design space for complex units like high-performance multipliers.
Introducing AC-Refiner: A New Approach to Circuit Optimization
A groundbreaking new framework called AC-Refiner proposes a novel solution by leveraging conditional diffusion models. This innovative approach redefines arithmetic circuit synthesis as a conditional image generation task. Imagine generating an image, but instead of a picture, you’re generating an optimized circuit design.
The core idea behind AC-Refiner is to guide a “denoising diffusion process” based on desired performance metrics, known as Quality-of-Results (QoRs). By doing so, AC-Refiner consistently produces high-quality circuit designs. Furthermore, the designs discovered during this process are used to fine-tune the diffusion model itself, allowing it to focus its exploration on the most promising design areas, closer to the ideal “Pareto frontier” of optimal designs.
How AC-Refiner Works
- Circuit Representation: Arithmetic circuits, specifically compressor trees and prefix adders, are transformed into image-like binary bitmaps. This allows the diffusion model, which is typically used for image generation, to process them.
- Model Training: The framework trains a diffusion model to predict and remove noise from these circuit “images” and also trains a neural cost predictor. This predictor estimates the real-world performance (delay and area) of a circuit design.
- Conditional Design Generation: This is where the magic happens. AC-Refiner uses the neural cost predictor to guide the diffusion process. It steers the generation towards designs that are predicted to have better QoRs. A “self-reflection” mechanism is also employed to ensure that while optimizing for performance, the generated designs remain structurally correct and valid.
- Design Legalization: After generation, a heuristic legalization procedure is applied to correct any minor design rule violations, ensuring the functional correctness of the circuits.
- Model Fine-Tuning: The high-quality designs discovered in each optimization round are used to fine-tune both the diffusion model and the cost predictor. This iterative process allows AC-Refiner to continuously improve its ability to find even better designs.
Also Read:
- Revolutionizing Hardware Design: How Agentic AI is Building Better Chips
- Generative AI Propels Robots to New Heights and Safer Landings at MIT CSAIL
Impressive Results
Experimental results demonstrate that AC-Refiner significantly outperforms existing state-of-the-art methods for multiplier optimization. It generates designs with superior Pareto optimality, meaning it achieves better trade-offs between performance (delay) and area. Compared to the best existing methods, AC-Refiner achieved up to a 15% reduction in delay and a 10% reduction in area. Its effectiveness was further validated by integrating the optimized multipliers into practical applications like systolic arrays, commonly used in AI accelerators, where it consistently showed performance advantages.
This innovative use of conditional diffusion models marks a significant step forward in automated arithmetic circuit design, promising more efficient and powerful digital systems for future technologies. You can read the full research paper for more technical details here: AC-Refiner: Efficient Arithmetic Circuit Optimization Using Conditional Diffusion Models.
