Power Modeling and Optimization for GPGPUs

Abstract

Modern graphics processing units (GPUs) supports tens of thousands of parallel threads and delivers remarkably high computing throughput. General-Purpose computing on GPUs (GPGPUs) is becoming the attractive platform for general-purpose applications that request high computational performance such as scientific computing, financial applications, medical data processing, and so on. However, GPGPUs is facing severe power challenge due to the increasing number of cores placed on a single chip with decreasing feature size. In order to explore the power optimization techniques in GPGPUs, I first build a power model for GPGPUs, which is able to estimate both dynamic and leakage power of major microarchitecture structures in GPGPUs. I then target on the power-hungry structures (e.g. register file) to explore the energy-efficient GPGPUs. In order to hide the long latency operations, GPGPUs employs the fine-grained multi-threading among numerous active threads, leading to the sizeable register files with massive power consumption. The conventional method to reduce dynamic power consumption is the supply voltage scaling. And the inter-bank tunneling FETs (TFETs) is the promising candidate compared to CMOS for low voltage operations regarding to both leakage and performance. However, always executing at the low voltage will result in significant performance degradation. In this study, I propose the hybrid CMOS-TFET based register file and allocate TFET-based registers to threads whose execution progress can be delayed to some degree to avoid the memory contentions with other threads to reduce both dynamic and leakage power, and the CMOS-based registers are still used for threads requiring normal execution speed. My experimental results show that the proposed technique achieves 30% energy (including both dynamic and leakage) reduction in register files with negligible performance degradation compared to the baseline case equipped with naive power optimization technique.