Dataflow dot product on networks of heterogeneous digit-serial arithmetic units

Abstract

(eng) In this paper we deal with a new high precision computation of the dot product. The key idea is to use hundreds of digit-serial arithmetic units that allow a massive digit-level pipelining. Parallel discrete-event simulations performed on a memory-distributed massively parallel computer show that with a limited number of arithmetic units, the computation of dot product when performed using a ``classical'' algorithmic technique (i.e. serial cumulative multiplications) is almost as fast as the case where an ``optimal'' divide-and-conquer algorithmic technique is used. Interconnection networks for both algorithmic techniques are considered.