Amortized Neural Optimization for Pre-Layout Signal Integrity Design Space Exploration using Differentiable Surrogates

This research presents an Amortized Neural Optimization (ANO) framework that attempts to solve the computational difficulties of pre-layout design space exploration (DSE) for high-speed signal integrity (SI) analysis. Most traditional electronic design automation (EDA) workflows use heavy simulations and iterative optimization algorithms, which are highly inefficient for multi-corner sweeps. Even if machine learning surrogate models speed up certain simulation steps, the optimization process still uses iterative black-box search methods. The ANO framework overcomes these challenges by using fully differentiable neural network surrogate models that allow analytical gradients to be extracted. These gradients are used to train a global optimization policy offline, which means the optimization process is significantly reduced. After

policy training is complete, the ANO policy can associate different channel contexts with design parameters that are optimum through a single, deterministic forward pass; this removes the need to use iterative black-box inference during the design process. The effectiveness and accuracy of this framework is showcased in three complex scenarios of SI design, which are DDR5 decision feedback equalization, 9-dimensional SerDes Tx/Rx co-equalization, and DDR3 DQS differential pair routing, which primarily target the optimization of eye diagram metrics.

Message: Hi FE, I need help with “Amortized Neural Optimization for Pre-Layout Signal Integrity Design Space Exploration using Differentiable Surrogates” in “Electrical Engineering”