COMPUTATIONAL RESEARCH in BOSTON and BEYOND (CRIBB)

Date	May 6, 2016
Speaker	Martin Herbordt Boston University
Topic	Scalable High Performance Computing with FPGA-based clusters: Is it (finally) time for FPGAs in HPC?
Abstract	The acquisition of Altera by Intel and the deployment by Microsoft of FPGAs in the Cloud indicate that widespread use of FPGAs for High Performance Computing may (finally) have arrived. This talk has two parts. In the first part we describe our work with FPGA-centric clusters with direct and programmable interconnects. These systems address fundamental limits on HPC performance by maximizing computational density, minimizing power, removing the bottleneck between processing and communication, and facilitating intelligent and application-aware processing in the network itself. For example, knowing the routing pattern a priori often enables congestion-free communication. We describe our 128-node publicly available cluster, Novo-G#, its software and IP infrastructure, and strong scaling of communication-bound applications such as long time-scale Molecular Dynamics. In the second part we discuss general issues such as prospects for FPGA-based HPC in the Cloud, programmability, integration with standard middleware, and potential applications.

Acknowledgements

We thank the generous support of MIT IS&T, CSAIL, and the Department of Mathematics for their support of this series.