Enhancing GPU-based Supercomputing through Workload and Communication Optimization
Steven S. Lumetta
Award year: 2009-2010
GPU-enabled clusters, which offer huge reductions in cost and incredible power efficiency relative to microprocessor-based systems, may provide a viable means to exascale computing. Before the wider science and engineering community can effectively harness this potential, however, we must find methods to overcome the complexity barriers imposed by their inherently hierarchical communication systems and heterogeneous processing resources. I believe that by building on several recent advances, we can generalize some of the successes demonstrated by our dedicated application experts on these clusters so as to make the performance advantages available to a broader segment of the community.