High Performance Computing (HPC) demands the highest performance technology combined with intricate programming to deliver efficient and effective solutions to the most challenging and complex computational problems. HPC systems are growing increasingly complex. Faults that occur during runtime are harder and harder to diagnose. Silent data corruptions, where bits get flipped and answers change without being detected at runtime, are more and more common.  We are researching technologies, algorithms and the feasibility of using probabilistic hardware to deliver 20x or more operations per Joule improvement to mission applications compared to commercial, non-probabilistic hardware. Accepting “faults”, willing to have imperfections, in exchange for more operations per joule allows users of HPC to get “good enough” answers to their problems far more quickly and efficiently than requiring “perfect” reliability

Adv Computing 2.jpg