Energy Analysis for Graphics Processors using Novel Methods & Efficient Multi-View Rendering

Sammanfattning: Real-time rendering is increasingly being performed on battery powered devices, such as laptops and mobile phones. As a result, performance in rendering time is not the only important factor, but the energy consumption is also becoming more and more important, as a lower energy usage directly translates to longer battery time. The main topic of this thesis is energy consumption of graphics processors. To examine this, we perform high-frequency measurements of a number of graph- ics devices’ power consumption. Simultaneously, we render real-time graphics workloads to analyze the power consumption for a number of devices, algorithms, and settings. We draw a number of conclusions for different platforms, e.g., that it is incorrect to assume a direct correlation between rendering time and per-frame energy consumption. We also present a method for evaluating if there is such a correlation on a specific, efficiently utilized, platform. This method uses Pareto frontiers to filter out measurements that are inefficient, with respect to render- ing time and energy consumption, and analyses only measured data points with a possible trade-off. Our long-term goal is to use our conclusions for developing energy-efficient algorithms, as well as raising awareness in the developer commu- nity to consider rendering time and energy consumption while developing real- time graphics algorithms. Furthermore, we develop and improve methods for cor- relating energy consumption on a per-frame basis with other information collected for specific frames, to enable improved analysis regarding real-time graphics en- ergy consumption. Our second topic is efficient multi-view rendering. We have developed an opti- mized algorithm for multi-view ray tracing, targeting auto-stereoscopic displays, which performs up to an order of magnitude faster than previous state of the art algorithms. In addition, we have examined the feasibility of enabling a proposed higher-dimensional rasterizer implemented in hardware to render multi-view and stereoscopic image sets in a single pass. We find it straightforward to adapt a higher-dimensional rasterizer to support multi-view rendering, and propose im- provements to enhance the rendering performance for such applications.