Distributed Power Management for 22nm AI Processor with Event-Driven Exponential Dual-Loop LDOs and Online Sparsity-Aware Droop Mitigation

A distributed power management solution is presented for a 22 nm digital compute-in-memory (DCIM) based AI processor featuring distributed dual-loop digital LDOs with event-driven exponential control for fast response, online learning-based sparsity-aware droop mitigation, and workload-driven current balance scheme among LDOs. The work demos the promising opportunity of deploying distributed management for robust and efficient AI computing, where the above techniques lead to $>71.6 \%$ droop reduction for diverse AI workloads and up to 33.3 % energy saving or 22.5 % performance improvement.