Bridging the Scalability Gap by Exploiting Error Tolerance for Emerging Applications

In recent years, there has been a surge in demand for intelligent applications. These emerging applications are powered by algorithms from domains such as computer vision, image processing, pattern recognition, and machine learning. Across these algorithms, there exist two key computational characteristics. First, the computational demands they place on computing infrastructure is large, with the potential to substantially outstrip existing compute resources. Second, they are necessarily resilient to errors due to their inputs and outputs being inherently noisy and imprecise.

Despite the staggering computational requirements and resilience of intelligent applications, current infrastructure uses conventional software and hardware methodologies. These systems needlessly consume resources for every bit of precision and arithmetic. To address this inefficiency and help bridge the performance gap caused by intelligent applications, this dissertation investigates exploiting error tolerance across the hardware-software stack. Specifically, we propose (1) statistical machinery to guarantee that accuracy is not compromised when removing work or precision, (2) a GPU optimization framework for work skipping and bottleneck mitigation, and (3) exploration of unconventional numerical representations to steer future hardware designs.