A Differentiable Simulator for Optimizing Time-Domain Analog CNN Accelerators

Abstract

Edge intelligence promises responsive, private, and energy-efficient sensing without continual dependence on remote compute. This demands convolutional neural network (CNN) accelerators that deliver substantially higher throughput and energy efficiency than conventional digital pipelines while preserving high accuracy. Time-domain analog CNN accelerators offer compact, high-throughput neural pipelines with reduced conversion and data-movement overhead, but also introduce modeling challenges. Hardware-aware models have been implemented in GPU-accelerated deep-learning libraries, but these are not designed for trajectory-dependent multiply-accumulate (MAC) operations where dynamic circuit behavior determines the final output. This work introduces a differentiable, GPU-accelerated CNN inference simulator that backpropagates through the analog accumulation trajectory. By making time-domain circuit dynamics compatible with automatic differentiation, the simulator allows gradient-based exploration of high-dimensional, heterogeneous hardware-parameter settings. Representative tuning studies show that layerwise tuning of time-dependent circuits can improve task-aware operating points rather than selecting a single global setting, and that gradient-based tuning outperforms a derivative-free baseline.