Better IC Designs in Less Time Don't be a Spice monkey → proper transistor-level dimensioning → going AI

Abstract: Sizing an analog or RF circuit is a huge nonlinear optimization problem. With at least three degrees of freedom per transistor, a 10-transistor circuit already spans a 30-dimensional design space: ten trial values per parameter give 1030 candidates. Symmetry and smart search algorithms help, but finding the optimum stays slow and costly. This is the realm of the Spice monkey. A good starting point shrinks the search dramatically. The classic routes - square-law/weak-inversion hand models with VGT, or EKV with gm/ID - are fast but inaccurate in modern CMOS. Seeding simulations with them still leads to slow tweak-and-simulate loops, local minima and hidden artifacts. ProMOST, developed some 25 years ago with Philips Research/ED&T, sizes and biases single transistors on model data from your own PDK, using its own solvers. Specify intent and boundaries (incl. VGT or gm/ID), sweep what remains, and read off the optimum: • Millisecond response on full BSIM/PSP-class models - no hand-model approximations. • Constrained sweeps (constant-ID / constant-gm) that encode design intent. • Noise, mismatch, distortion, S-parameters and IIP3 at every bias point. • W-L design-space maps with spec filters, constrained optima and Pareto fronts. • Optional: bi-directional Cadence link and Python/MATLAB/Jupyter scripting. A 10-transistor circuit thus becomes ten 3-dimensional problems, with all higher-order effects included. This quickly gives a good starting point for short simulator optimization loops, and shows early whether the specs University of Ulm | Institute of Microelectronics | 89081 Ulm Seite 2 von 2 are reachable with your topology, assumptions and PDK. You are in charge of the design, with limited Spice monkeying. Still easily missed: frequency-dependent distortion, large-signal excursions, impulse sensitivity, proper use of small-signal parameters. E.g. ProMOST's Volterra outputs can capture much of this, but doing so by hand is laborious. The last part of the talk therefore goes AI: gluing tools to an AI engine that builds on ProMOST's intentdriven output, finds roadblocks to the target specs and automates the final optimization. NDA/IP issues are briefly addressed. The AI is the Spice monkey; you are the monkey master. The goal: better designs in less time - and staying in command of your design.