In this video, Veljko Milutinovic and Oliver Pell of Maxeler present: DataFlow Computing for Exascale HPC. Recorded a the HPC Advisory Council Spain Workshop 2012.
Data movement and energy are becoming the biggest constraints in HPC. DataFlow computers have been shown to provide order of magnitude improvements in space and power consumption by focusing first on optimizing data movement in a computer system, then utilizing massive parallelism between thousands of tiny dataflow cores. In this talk, we will introduce dataflow computing and outline the advantages of the approach as well as what kinds of operations map most effectively to the dataflow paradigm. We will discuss several case studies of scientific applications running on different heterogeneous control-flow / dataflow machines, including clusters of heterogeneous compute nodes where each node has control flow CPUs and dataflow engines, and loosely-coupled systems where control-flow nodes utilize dataflow engines as network resources over InfiniBand.