SIMD made explicit

Abstract

Low energy consumption has become one of the most important topics in computing. With single CPUs consuming as much as 115 Watt, engineers have been looking for ways to reduce energy consumption while maintaining high computational performance. Often wide SIMD architectures are used to achieve this, exploiting data parallelism to keep the required clock frequency low for a given compute constraint. In this paper, we propose a wide SIMD architecture with explicit datapath to further optimize energy efficiency without sacrificing computation power. To have a detailed comparison, both the proposed wide SIMD architecture and its transparent bypassing counterpart are implemented in HDL and synthesized with a TSMC 40nm low power library. The power estimation is derived from actual toggle rates generated by post-synthesis simulation. Our experimental results show that with explicit bypassing the overall energy consumption can be reduced up to 44% compared to the corresponding SIMD architecture with transparent bypassing.