Improving POF Quality in Multi Objective Optimization of Analog ICs via Deep Learning

Cakici T. O., Islamoglu G., Guzelhan S. N., Afacan E., DÜNDAR G.

24th IEEE European Conference on Circuit Theory and Design, ECCTD 2020, Sofija, Bulgaristan, 7 - 10 Eylül 2020, (Tam Metin Bildiri) identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/ecctd49232.2020.9218272
  • Basıldığı Şehir: Sofija
  • Basıldığı Ülke: Bulgaristan
  • Boğaziçi Üniversitesi Adresli: Evet

Özet

Multi-objective optimization (MOO) is commonly used in analog circuits to reveal the trade-offs among design specifications via Pareto optimal fronts (POF). Although the general trend of POF can be found in a reasonable time with MOO, a high-quality POF requires an excessive number of iterations, which results in extremely long synthesis times. In this paper, single-objective optimization (SOO) is utilized to increase the POF quality rather than running MOO algorithms for long duration. Moreover, deep neural networks (DNN) are used to replace SPICE, which reduces the synthesis time further. This approach provides up to 50.62% improvement in POF quality and DNNs speed up the process up to 29.6x.