Authors: Brady Taylor, J. Darby Smith, Hemanth Kolla, and Michael J. Schmidt (Sandia National Laboratories)
Abstract: Graphics processing units (GPUs) excel at workloads dependent on linear algebra, covering a wide range of topics like artificial intelligence and quantum mechanics. However, in workloads that are not well captured by linear algebra, such as probabilistic systems, GPUs may not provide optimal performance.
In this work, we demonstrate a neuromorphic advantage in energy efficiency for simulating turbulent flows with stochastic methods. We observe energy savings across the board and runtime speed-up for some methods when comparing a neuromorphic platform to equivalent workloads on GPU and CPU. Specifically, our neuromorphic implementation achieves 1.52x speed-up and 37.46x less energy consumption, compared to a GPU running the same 1.5 million random walks over discretized space. For 100,000 walks in continuous space, GPU outperforms neuromorphic in execution time but consumes 2.37x more energy. As the technological landscape for neuromorphic systems rapidly evolves, their performance relative to current hardware is likely to significantly improve.
Presentation: file