New Views for Stochastic Computing: From Time-Encoding to Deterministic Processing

Abstract

Stochastic computing (SC), a paradigm first introduced in the 1960s, has received considerable attention in recent years as a potential paradigm for emerging technologies and ''post-CMOS'' computing. Logical computation is performed on random bitstreams where the signal value is encoded by the probability of obtaining a one versus a zero. This unconventional representation of data offers some intriguing advantages over conventional weighted binary. Implementing complex functions with simple hardware (e.g., multiplication using a single AND gate), tolerating soft errors (i.e., bit flips), and progressive precision are the primary advantages of SC. The obvious disadvantage, however, is latency. A stochastic representation is exponentially longer than conventional binary radix. Long latencies translate into high energy consumption, often higher than that of their binary counterpart. Generating bit streams is also costly. Factoring in the cost of the bit-stream generators, the overall hardware cost of an SC implementation is often comparable to a conventional binary implementation. This dissertation begins by proposing a highly unorthodox idea: performing computation with digital constructs on time-encoded analog signals. We introduce a new, energy-efficient, high-performance, and much less costly approach for SC using time-encoded pulse signals. We explore the design and implementation of arithmetic operations on time-encoded data and discuss the advantages, challenges, and potential applications. Experimental results on image processing applications show up to 99% performance speedup, 98% saving in energy dissipation, and 40% area reduction compared to prior stochastic implementations. We further introduce a low-cost approach for synthesizing sorting network circuits based on deterministic unary bit-streams. Synthesis results show more than 90% area and power savings compared to the costs of the conventional binary implementation. Time-based encoding of data is then exploited for fast and energy-efficient processing of data with the developed sorting circuits. Poor progressive precision is the main challenge with the recently developed deterministic methods of SC. We propose a high-quality down-sampling method which significantly improves the processing time and the energy consumption of these deterministic methods by pseudo-randomizing bitstreams. We also propose two novel deterministic methods of processing bitstreams by using low-discrepancy sequences. We further introduce a new advantage to SC paradigm-the skew tolerance of SC circuits. We exploit this advantage in developing polysynchronous clocking, a design strategy for optimizing the clock distribution network of SC systems. Finally, as the first study of its kind to the best of our knowledge, we rethink the memory system design for SC. We propose a seamless stochastic system, StochMem, which features analog memory to trade the energy and area overhead of data conversion for computation accuracy.