NVIDIA Explores Generative AI Models for Improved Circuit Design

Generative models have made significant progress in recent years, from large-scale language models (LLMs) to creative image and video generation tools. According to the NVIDIA Technical Blog, NVIDIA is now aiming to apply these advances to circuit design to improve efficiency and performance.

Complexity of circuit design

Circuit design presents a challenging optimization problem. Designers must balance multiple conflicting objectives, such as power consumption and area, while meeting constraints such as timing requirements. The design space is vast and combinatorial, making it difficult to find optimal solutions. Existing methods have relied on handcrafted heuristics and reinforcement learning to navigate this complexity, but these approaches are computationally intensive and often lack generalization.

Introduction to CircuitVAE

In a recent paper CircuitVAE: Efficient and Scalable Latent Circuit Optimization, NVIDIA demonstrates the potential of Variational Autoencoders (VAEs) in circuit design. VAEs are a class of generative models that can produce better prefix adder designs at a fraction of the computational cost of previous methods. CircuitVAE embeds a computational graph in a continuous space and optimizes a surrogate of a physical simulation learned via gradient descent.

How CircuitVAE Works

The CircuitVAE algorithm involves embedding circuits in a continuous latent space and learning a model that predicts quality metrics such as area and delay from this representation. This cost prediction model, instantiated as a neural network, allows gradient descent optimization in the latent space, thereby bypassing the task of combinatorial search.

Training and Optimization

The training loss of CircuitVAE consists of the standard VAE reconstruction and regularization loss and the mean squared error between the real and predicted domains and lags. This dual loss structure facilitates gradient-based optimization by organizing the latent space according to the cost metric. The optimization process involves selecting latent vectors using cost-weighted sampling and refining them through gradient descent to minimize the cost estimated by the prediction model. The final vectors are then decoded and synthesized into a prefix tree to evaluate the real cost.

Results and Impact

NVIDIA tested CircuitVAE on circuits with 32 and 64 inputs using the open-source Nangate45 cell library for physical synthesis. The results, shown in Figure 4, demonstrate that CircuitVAE consistently achieves lower cost than baseline approaches, thanks to its efficient gradient-based optimization. On practical tasks involving proprietary cell libraries, CircuitVAE outperforms commercial tools, demonstrating better Pareto frontiers for area and delay.

Future outlook

CircuitVAE demonstrates the transformative potential of generative models in circuit design by shifting the optimization process from discrete space to continuous space. This approach significantly reduces computational costs and offers hope for other hardware design areas such as place and route. As generative models continue to evolve, they are expected to play an increasingly central role in hardware design.

For more information on CircuitVAE, visit the NVIDIA Technology Blog.

Image source: Shutterstock