Deep Learning Autoencoder-based Compression for Current Source Model Waveforms

Abstract

Modeling complex cell behavior is critical for accurate static timing analysis. Huge waveform data needed for current source models explodes technology file size and degrades design flow performance. We used deep learning nonlinear Autoencoders to compress voltage and current waveforms and compared them with singular component analysis approach. Autoencoders gave up to 3.37x compression ratio for voltage waveforms with average percentage error below 0.85% better than SVD approach. Autoencoders achieved 1.7x compression ratio for complex rising-edge current waveforms at model loss of 7.6e-5 and comparable results to SVD approach for the falling-edge waveforms. SVD remains more computationally efficient than Autoencoders.