EREER: Energy-aware register file and execution unit using exploiting redundancy in GPGPUs

Abstract

Nowadays, the use of GPGPUs is growing in high-performance computing including embedded system. Demanding more processing power increase the size of Register File (RF) and Execution Unit (EU) in GPGPUs, that increase the power consumption. However, energy and power consumption are vital for the embedded system due to using a battery and a simple cooling system. In this paper, initially, we have proposed a simple method to identify duplicated data in RF of GPGPUs. Afterward, we propose a compression method to improve the energy efficiency of RF by eliminating duplicated data and consequently unallocating some of RF banks. Experimental results on standard benchmarks show that our compression method reduces the total RF power consumption by 15% on average by considering overhead degradation. Furthermore, we propose a computation reuse method in the EU to exploit computation redundancy. This method utilizes the compression information of RF to identify the identical computations and turn off the processing cores that execute them. Moreover, our computation reuse method improves the EU energy efficiency by 28.8% on average.