Approximate signed binary integer multipliers for arithmetic data value speculation

Abstract

Arithmetic data value speculation increases the throughput of a processor pipeline by speculatively issuing the dependent operations of an arithmetic operation based on the early arrival of an approximate result. Suitable approximate multipliers calculate a product faster than an exact multiplier, with an associated probability that the approximate product is correct. This paper presents the design of a family of signed approximate multipliers for use in a speculative data path. A signed 32x32 bit multiplier synthesised with the TSMC Artisan 180nm SAGE-X™ cell library is found to be 20% faster than a full-adder based tree multiplier, with a probability of error less than 14% for benchmark applications.