A workload-aware neuromorphic controller for dynamic power and thermal management

Abstract

A workload-aware low-power neuromorphic controller for dynamic power and thermal management in VLSI systems is presented. The neuromorphic controller predicts future workload and temperature values based on the past values and preemptively regulates supply voltage and frequency. Our specific contributions include: (1) implementation of a digital and analog version of the controller in 45nm CMOS technology, resulting in 3% performance hit with a power overhead in the range of 10–150 microwatts from the controller circuit, (2) higher prediction accuracy compared to a software based OS-governed DVFS scheme, reducing wasted power and improving error margins, (3) power savings of up to 52% and improvement of up to 15% compared to the OS based scheme, (4) implementing DVFS when the processor is memory bandwidth limited for additional savings and (4) accurate temperature prediction to proactively implement DVFS and prevent CPU shutdown.The digital design has minimal power overhead and is more reconfigurable, while analog design is better suited for nonlinear and complex computational tasks.