UC Berkeley

Improving the energy efficiency of processor systems-on-chip (SoCs) is key to improving the performance and utility of thermally-limited servers, battery-constrained mobile devices, and energy-harvesting Internet-of-Things nodes. Adaptive voltage scaling adjusts voltage levels at runtime in response to changes in workload, reducing voltage when additional performance is not required. The granularity of this voltage scaling greatly impacts its effectiveness. Fine-grained adaptive voltage scaling (FG-AVS) in time reduces wasted energy by tracking microsecond-scale changes in workload; FG-AVS in space independently adjusts voltage levels in many different portions of an SoC, saving additional energy. FG-AVS has the potential to save considerable energy, but it has not yet seen adoption in commercial systems due to several challenges that complicate its implementation. Integrated voltage regulators are required to supply the many voltages required and switch quickly between modes, but efficient implementations impose high area overhead. Clocking, synchronization, and power management have their own requirements and difficulties.

This work presents key components of fully-featured SoCs that enable demonstration of FG-AVS. Integrated simultaneous-switching switched capacitor voltage regulators are presented that achieve high conversion efficiency when coupled with an adaptive clock generator. Power management is accomplished with programs run on a dedicated power management unit (PMU), and a pausible bisynchronous FIFO circuit that can achieve low-latency communication between asynchronous voltage domains is described. These components are integrated into systems that demonstrate the potential energy savings of FG-AVS. The Raven-3 testchip demonstrates efficient voltage regulation supplying a complicated digital load; the system achieves 26.2 GFLOPS/W while operating its processor under the generated supply voltage and adaptive clock. The Raven-4 testchip includes integrated power management circuits that allow fully integrated feedback for FG-AVS. The programmable PMU can run a wide variety of power management algorithms, including an AVS algorithm that saves 39.6% energy in a synthetic microbenchmark with minimal performance penalty. The Hurricane-1 testchip implements multiple independent voltage domains and hardware counters that can be used for power management. The Hurricane-2 testchip features finer spatial partitioning and more effective instrumentation for power management; simulation results show up to 13.3% energy savings for an algorithm exercising FG-AVS in time and 46.0% energy savings for an algorithm using FG-AVS in space. Together, these testchip implementations show the potential of FG-AVS to save energy in production SoCs.