Details

Autor(en) / Beteiligte

• Chen, Te-Hsuan
• Ting, Paishun
• Hayes, John P.

Titel

Achieving progressive precision in stochastic computing

Ist Teil von

• 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), 2017, p.1320-1324

Ort / Verlag

IEEE

Erscheinungsjahr

2017

Quelle

IEEE Xplore

Beschreibungen/Notizen

• Stochastic computing (SC) performs approximate arithmetic with variable-length probabilistic bit-streams. Its potential advantages include low hardware cost, low power, and variable precision. Increasing the bit-stream length N generally improves precision, but not necessarily in a consistent or predictable way due to the randomness of the bit patterns. A desirable property for SC is progressive precision (PP), meaning that initial parts of a bit-stream X provide steadily improving estimates of X's exact value. In most stochastic circuits, PP is either absent or, if present, based on the vague notion that "longer is better." We propose a concept called accurate truncated progressive precision (ATPP) which provides PP that has a well-defined accuracy guarantee for all 2'-bit initial subsequences of a stochastic number. We then describe a new stochastic number generator CAPE that, when used to generate inputs for a combinational stochastic circuit, guarantees ATPP for the circuit's output bit-streams. Experimental results are presented which confirm CAPE's ATPP property.