Beyond ideal DVFS through ultra-fine grain Vdd-hopping

Abstract

DVFS is the de-facto standard for low energy Multi-Processor SoCs. It is based on the simple, yet efficient principle of lowering the supply voltage (Vdd) to the minimum threshold that satisfies the frequency constraint (fclk) required by the actual workload. An ideal-DVFS deals with the availability of on-chip high resolution voltage regulators that can deliver the supply voltage with a fine step resolution, a design option that is too costly. While previous research focused on alternative solutions that can achieve, or at least get close to, the efficiency of ideal-DVFS while using a discrete set of supply voltages, this work introduces Ultra-Fine Grain Vdd-Hopping (FINE-VH), a practical methodology that brings DVFS beyond its theoretical limit. FINE-VH leverages the working principle of Vdd-Hopping applied within-the-core by means of a layout-assisted, level-shifter free, dynamic dual-Vdd control strategy in which leakage currents are minimized through an optimal timing-driven poly-bias assignment procedure. We propose a dedicated back-end flow that guarantees design convergence with minimum area/delay overhead for a cutting-edge industrial Fully-Depleted SOI (FDSOI) CMOS technology at 28 nm. Experimental results demonstrate FINE-VH allows substantial power savings w.r.t. coarse-grain (i) ideal-DVFS, (ii) Vdd-Hopping, (iii) Vdd-Dithering, when applied on the design of a RISC-V architecture. A quantitative analysis provides an accurate assessment of both savings and overheads while exploring different design options and different voltage settings.