Design of Ultra-Low Power Scalable-Throughput Many-Core DSP Applications

Abstract

We propose a system-level solution in designing process variation aware (PVA) scalable-throughput manycore systems for energy constrained applications. In our proposed methodology, we leverage the benefits of voltage scaling for obtaining energy efficiency while compensating for the loss in throughput by exploiting parallelism present in various DSP designs. We demonstrate that such a hybrid method consumes 6.27%-28.15% less power as compared to simple dynamic voltage scaling over different workload environments. Design details of a prototype chip fabricated on 90nm technology node and its findings are presented. Chip consists of 8 homogeneous FIR cores, which are capable of running from near-threshold to nominal voltages. In our 20 chip population, we observe 7% variation in speed among the cores at nominal voltage (0.9V) and 26% at near threshold voltage (0.55V). We also observe 54% variation in power consumption of the cores. For any desired throughput, the optimum number of cores and their optimum operating voltage is chosen based on the speed and power characteristics of the cores present inside the chip. We will also present analysis on energy-efficiency of such systems based on changes in ambient temperature.