Cross-Domain Optimization of Low-Power Mixed-Signal Sensor Systems Under Classification Accuracy Constraints

Optimizing mixed-signal systems-on-chips (SoCs) is a challenging task, especially when they involve both analog building blocks and machine-learning (ML) algorithms which make the overall system performance hard to predict. This article proposes a methodology for optimizing complex mixed-signal sensor systems, which consists of three steps: 1) modeling Pareto-optimal local performance of individual circuits; 2) building a macromodel of the mixed-signal system performance, including local nonidealities; and 3) finding the optimal set of system-level parameters using numerical optimization. The methodology is applied to the use case of an electrocardiogram (ECG) sensor system with embedded arrhythmia classification, with the objective of minimizing the system power consumption while imposing constraints on inference accuracy. Different numerical optimization methods are compared to evaluate their respective performance on the minimization task: gradient-based, genetic, or Bayesian optimization. The proposed methodology achieves the optimization in a reasonable execution time of 10 h, which had not been previously demonstrated on such complex mixed-signal system. Among the different methods tested, Bayesian optimization provides the best-success rate of 97% and fastest convergence speed.