Optimizing the Energy Efficiency of Switched-Capacitor Converters in Multiprocessor System-on-Chips with a Preset DVFS Policy

Abstract

Some multiprocessor system-on-chips (MPSoCs) provide designers with workload information in early chip planning stage to optimize the chip's performance. In our work, we focus on the energy efficiency of switched-capacitor converters (SCCs) in MPSoCs to efficiently support dynamic voltage and frequency scaling (DVFS). Two things motivate us to propose our design flow in MPSoCs, i.e. 1) Recent work allocates Metal-Insulator-Metal (MIM) capacitance and selects a converter ratio for each SCC to reach a higher energy efficiency. However, it brings the high overhead of its hardware implementation for various voltage and current demands. 2) DVFS policy has an unbalance distribution of its scaling decisions, which further results in an unbalance distribution of the voltage and current demands. Based on these, in our design flow, the hotspot recognition and the look-up table techniques help us to overcome the shortcomings of the recent work and harvest the benefits of the recent work. The experiment explores the different hotspot capacities' effects on energy efficiency and overhead.