Today, power consumption is as important as speed in almost all electronic design. Handheld electronic devices, such as cellular phones, personal digital assistants, and handheld game consoles have become commonplace and consumers are demanding longer battery lifetimes without additional weight, in addition to greater functionality and performance. This research addresses this critical need for developing energy-efficient graphics hardware for use in handheld devices.
The key objective of this research is to explore the use of asynchronous (or clockless) logic to reduce power consumption in the graphics processing units (GPUs) of handheld devices, while maintaining a given level of performance. An asynchronous graphics processor promises to reduce power requirements in many ways. Whereas the constantly running clock in synchronous logic causes circuits to consume power even when no work is being done, clockless logic is essentially powered down when idle. Thus, the asynchronous graphics pipeline would adjust naturally to the constantly varying demands of graphics. The graphics problem is also a good match with asynchronous design because it has many pipeline stages, is tolerant to latency, and can take advantage of parallelism at a fine granularity everywhere in the pipeline. Furthermore, other power benefits specific to graphics pipelines are expected to be achieved by introducing two new capabilities that are greatly facilitated by asynchronous design: on-demand and variable-precision computation.
As mobile devices become part of our daily lives, it has become more and more important to reduce energy consumption in order to lengthen battery life. 3D graphics, used not only for entertainment but also increasingly in augmented reality, consumes a great deal of energy. Further, many 3D applications that have so far been available only on mainframe and desktop computers are gradually being adapted to tablet devices, including mapping of urban environments, computer-aided design, photorealistic rendering, physics-based simulation, and 3D sketches and art, to name a few. All of these applications are expected to present a challenge to energy efficiency. In this project, we have tackled that problem. We carefully studied the two main ways in which graphics processors consume energy: computation and communications. We discovered that by reducing the precision of the computation, the number of bits used, we could save up to 70% of the energy used to perform arithmetic, without reducing the quality of the resulting images. In order to reduce communications energy, we designed a new, general-purpose data compression scheme that has resulted in savings ranging from 34% to 65% of the memory bandwidth. This should result not only in lower energy consumption, but also promises to improve performance because of fewer memory lookups required. To conduct this research, we developed a model of the energy consumed in graphics processors. To validate the model, we compared it to a hardware graphics processor, and obtained results accurate to within 5%. All of our techniques have been successfully tested on real-world applications. The educational impact has been substantial. Several students have learned research methods while working on this project. One has earned a PhD and has gone on to a career in industrial research. Another student earned an undergraduate degree, and decided to continue for an advanced degree.
