Cooperative high-performance computing with FPGAs - matrix multiply case-study

• Main Author: Munafo, Robert
• Other Authors: Herbordt, Martin C.
• Language: en_US
• Published: 2018
• Subjects: Computer engineering
• Online Access: https://hdl.handle.net/2144/30740

In high-performance computing, there is great opportunity for systems that use FPGAs to handle communication while also performing computation on data in transit in an ``altruistic'' manner--that is, using resources for computation that might otherwise be used for communication, and in a way that improves overall system performance and efficiency. We provide a specific definition of \textbf{Computing in the Network} that captures this opportunity. We then outline some overall requirements and guidelines for cooperative computing that include this ability, and make suggestions for specific computing capabilities to be added to the networking hardware in a system. We then explore some algorithms running on a network so equipped for a few specific computing tasks: dense matrix multiplication, sparse matrix transposition and sparse matrix multiplication. In the first instance we give limits of problem size and estimates of performance that should be attainable with present-day FPGA hardware.