Neuro-Spiking-Based Nonlinear Time-Optimal Control for VRM

Voltage Regulator Module (VRM) applications are experiencing rapidly increasing power levels driven by the growth of AI-centric data centers. Concurrently, transient performance requirements have become more stringent, with current slew rates exceeding several amperes per nanosecond. This paper introduces a novel control technique that achieves transient responses beyond 1 A/ns without increasing the switching frequency of the power stage. The controller, fabricated in a 28 nm CMOS process, combines a conventional steady-state loop with a fully analog spiking neural network trained through machine learning to approximate time-optimal control. The approach is validated on a 16-phase TLVR operating at 7 V input and 0.7 V/1000 A output, demonstrating that fast transient responses can be attained at high input voltages and low switching frequencies, thereby improving efficiency and simplifying thermal management.