Reverse engineering power management on NVIDIA GPUs - A detailed overview

Abstract

Power management in Open Source operating systems is currently limited by the fact that hardware companies do not release enough documentation to write the most power-efficient drivers. This problem is even more present in GPUs despite having the highest performance-per-Watt ratio found in today's processors. This paper presents an overview of GPUs from a power management point of view and what power management features have been found when reverse engineering modern NVIDIA GPUs. This paper finally discusses about the possibility of achieving self-management on NVIDIA GPUs and also discusses the future challenges that will be faced by the community when autonomic systems like Broadcom's videocore will become mainstream.