A Native Stochastic Computing Architecture Enabled by Memristors

Abstract

A two-terminal memristor device is a promising digital memory for its high integration density, substantially lower energy consumption compared to CMOS, and scalability below 10 nm. However, a nanoscale memristor is an inherently stochastic device, and extra energy and latency are required to make a deterministic memory based on memristors. Instead of enforcing deterministic storage, we take advantage of the nondeterministic memory for native stochastic computing, where the randomness required by stochastic computing is intrinsic to the devices without resorting to expensive stochastic number generation. This native stochastic computing system can be implemented as a hybrid integration of memristor memory and simple CMOS stochastic computing circuits. We use an approach called group write to program memristor memory cells in arrays to generate random bit streams for stochastic computing. Three methods are proposed to program memristors using stochastic bit streams and compensate for the nonlinear memristor write function: voltage predistortion, parallel single-pulse write, and downscaled write and upscaled read. To evaluate these technical approaches, we show by simulation a memristor-based stochastic processor for gradient descent optimization, and k-means clustering. The native stochastic computing based on memristors demonstrates key advantages in energy and speed in compute-intensive, data-intensive, and probabilistic applications.

Introduction to stochastic computing and its challenges

“…Some of the nanoscale semiconductor technologies that are being investigated as possible replacements for CMOS circuits (in light of the anticipated demise of Moore's Law) seem well-suited to SC. These include memristors, which have been called inherently stochastic devices [11]. The success of these nanotechnologies in the construction of practical stochastic circuits remains to be seen, however.…”

On the Functions Realized by Stochastic Computing Circuits

“…It is also possible to combine non-random and pseudo-random bit-streams, but the results have been unpromising. "True" random sources, made possible by nanotechnologies like memristors [16] and magnetic-tunnel junction devices [20] have also been proposed recently for SC.…”

“…More recently, new applications have appeared that involve probabilistic or error-tolerance issues for which SC is well suited, such as image processing [2] [17], simulation of probabilistic systems [9] [21], data recognition and mining [11], and decoders for channel codes ranging from LDPC to polar codes [13] [28]. Furthermore, novel physical technologies are emerging such as memristors that have native stochastic features [16]. Despite these successes, many gaps exist in our understanding of SC and its potential applications.…”

Analog-to-stochastic converter using magnetic-tunnel junction devices

“…To take advantage of the area efficiency on stochastic computation, the signal-conversion overhead needs to be mitigated. In [10], a concept of the analog-to-stochastic conversion has been proposed using memristors. However, the switching behaviour of the memristor is very slow (the order of ms).…”