An Area- and Power-Efficient Stochastic Number Generator for Bayesian Sensor Fusion Circuits

Abstract

In recent years, stochastic computing became popular in Bayesian circuits implementation because it enables compact and low power architectures. These architectures use Stochastic Number Generators (SNGs) that encode data in random bit-streams. SNGs are composed of Random Number Generators (RNGs) which contribute significantly to the circuit area and power consumption: according to our measurements, up to 29% of the area and 85% of the circuit power consumption, excluding memories. In this paper, we compare SNG implementations in terms of accuracy, area and power consumption. Furthermore, we propose a new SNG architecture that uses a single RNG for the whole design in order to generate the required stochastic bit-streams. The proposed architecture allows to save up to 11% of area and 58% of power consumption compared to the state of the art, with no significant accuracy loss.

An Energy Efficient Multi-Rail Architecture for Stochastic Computing: A Bayesian Sensor Fusion Case Study

“…2 shows the architecture of our Bayesian sensor fusion circuit. The independent random numbers are given by the shift register, as proposed in [6] and developed in the next subsection.…”

“…The second part, the Probability Encoder (PE), replaces the standard comparator in each cell of the matrix and, since it is smaller, this solution saves area and power consumption. Taking advantage of these two works, in [6], Belot et al propose a Shifted Register Isolator (SRI), which does not place the isolation at the bit-stream level anymore, but at the random number level. This becomes more interesting than Chen et al's isolators [7] when the size of the random numbers is smaller than the number of independent computations (number of rows here).…”

“…Moreover, one can notice that the architecture is also slightly different from the one presented in [6]. We indeed optimize it with additional registers just after the WG in order to remove glitches occurring in PEs and thus reduce the power consumption.…”

“…Bayesian inference appears to be a good way to build fault-tolerant and explainable models for sensor fusion. Recently, architectures based on Stochastic Computing (SC), a non-standard representation introduced in 1956 by Von Neumann [1] and later developed by Gaines [2], have been successfully used to implement sensor fusion models and temporal filters [3]- [6].…”

“…These SNGs are responsible of the major part of SC circuit power consumption and area. Several works have successfully found ways to reduce this impact [6]- [8], but this is not the main purpose of this paper.…”