A compact, fast, energy-efficient, and scalable stochastic dot-product circuit.
There have been many demonstrations of efficient hardware for dot product computation, which is the most common operation in stochastic neural networks. In addition, there are numerous reports of stochastic neuron, a unique feature of Boltzmann machines. An efficient and scalable hardware solution combining both functionalities could increase efficiency and performance.
Researchers at the University of California, Santa Barbara have developed a compact, fast, energy-efficient, and scalable stochastic dot-product circuit. This circuit is based on two types of memory devices – metal oxide memristors and embedded floating gate memories. Through mixed-signal implementation, efficient stochastic operation is achieved by utilizing circuit’s noise, which can be intrinsic and/or extrinsic to the array of memory cells. This stochastic dot-product circuit is useful for many applications such as solving optimization problems and developing probabilistic neural networks.