Limits to semiconductor scaling have led to innovative processor designs, such as general-purpose GPUs and, more recently, mixed-precision processor support for artificial intelligence deep-learning workloads. Even more exotic processor features are expected soon, requiring domain scientists, computational scientists and application developers to modify and sometimes fundamentally rethink their algorithms and applications. In this talk we describe a new data-analytics application, CoMet, for solving problems in computational genomics, with uses such as biofuels research and the discovery of genetic causes of such conditions as Alzheimer’s disease and opioid addiction. By using the mixed-precision tensor-core hardware on Oak Ridge National Laboratory’s Summit system, we have achieved 20,000 times to 300,000 times improvement over the previous state of the art and have reached 2.36 exaops of performance on Summit, the first science application to break the mixed-precision exascale barrier. In this talk we describe the methods used to achieve high performance as well as results from computational experiments with CoMet.
