DVFS using clock scheduling for Multicore Systems-on-Chip and Networks-on-Chip

Abstract

A modern System-on-Chip (SoC) contains processor cores, application-specific process- ing elements, memory, peripherals, all connected with a high-bandwidth and low-latency Network-on-Chip (NoC). The downside of such very high level of integration and con- nectivity is the high power consumption. In CMOS technology this is made of a dynamic and a static component. To reduce the dynamic component, Dynamic voltage and Fre- quency Scaling (DVFS) has been adopted. Although DVFS is very effective chip-wide, the power optimization of complex SoCs calls for a finer grain application of DVFS. Ideally all the main components of an SoC should be provided with a DVFS controller. An SoC with a DVFS controller per component with individual DC-DC converters and PLL/DLL circuits cannot scale in size to hundreds of components, which are in the research agenda. We present an alternative that will permit such scaling. It is possible to achieve results close to an optimum DVFS by hopping between few voltage levels and by an innovative application of clock-gating that we term as clock scheduling. We obtain an effective clock frequency by periodically killing some clock cycles of a master clock. We can apply voltage scaling for some of the periodic clock schedules which yield effective clock 1/2, 1/3, . . . By dithering between few voltages we obtain results close to an ideal DVFS system in simple pipelined circuits and in a complex example, a NoC's switch. Again in the context of a NoC, we show how clock scheduling and voltage scaling can be automatically determined by means of a proportional-integral loop controller that keeps track of the network load. We describe in detail its implementation and all the circuit-level issues that we found. For a single switch, result shows an advantage of up to 2X over simple frequency scaling without voltage scaling. By providing each NoC's switch with our simple DVFS controller, power saving at network level can be significantly more than what a a global DVFS controller can get. In a realistic scenario represented by network traces generated by video applications (MPEG, PIP, MWD, VoPD), we obtain an average power saving of 33%. To reduce static power, the Power-Gating (PG) technique is used and consists in switching- off power supply of unused blocks via pMOS headers or nMOS footers in series with such blocks. Even though research has been done in this field, the application of PG to NoCs has not been fully investigated. We show that it is possible to apply PG to the input buffers of a NoC switch. Their leakage power contributes about 40-50% of total NoC power, hence reducing such contribution is worthwhile. We partitioned buffers in banks and apply PG only to inactive banks. With our technique, it is possible to save about 40% in leakage power, without impact on performance