Exploration of Linear-Algebra Graph Algorithms on the Graphcore IPU

Abstract

More specialized hardware is being created to improve the efficiency of demanding algorithms. Graphcore introduced their Intelligence processors: a specialized true MIMD architecture which specializes in machine learning problems. Earlier work has shown that this architecture can also be used to solve breadth-first search competitively. We explore the possibilities of this machine by implementing several graph algorithms on this machine. Specifically, we look at Sparse-matrix vector multiplication and its usage within an algebraic approach to the breadth-first search algorithm, and show an algebraic solution to Prim's algorithm on the machine. Finally, we discuss the possibilities and problems of the Graphcore IPU.