APEX: Recommending Design Flow Parameters Using a Variational Autoencoder

Modern physical synthesis flows employ thousands of parameters. Choosing the optimal set is an important problem. We introduce an offline ML based system that suggests parameter values to improve congestion, timing, and power. We employ a variational autoencoder to encode a high dimensional space into a low dimensional latent space from physical synthesis jobs and generate parameters and values. This compressed information requires minutes to predict instead of multiple days and large machine resources required for existing auto-tuner methods. Our results show this method is over 10 times faster than the present state of the art with similar quality.