GPU design in a power-limited era
Brucek Khailany · 2013
As we enter this era of power-limited computing with GPUs playing a leading role in parallel processor architecture, interesting design challenges need to be faced and new tradeoffs need to be considered. For example, aggressively reducing operating voltage closer to threshold has been proposed as one way to overcome power limitations. However, this approach can introduce complications related to variation-related effects and yield issues, power delivery losses at low voltage, and significant degradations in performance per mm2. Alternative approaches to improving power efficiency include micro-architectural techniques for eliminating wasteful energy in data and instruction operand delivery on-chip, particularly in the presence of data parallelism. Many of these and other techniques will play a key part in continuing to scale GPU performance in the presence of power dissipation limitations in future technologies.