Flash based in-memory multiply-accumulate realisation: A theoretical study

In memory computing is gaining traction as a technique to implement the Multiply Accumulate (MAC) operation on edge network devices, to perform neural network inference while reducing energy expended in memory-fetch. The voltage developed along a bit-line is an analog representation of the MAC value and needs to be digitized for further processing. In this paper we propose to use the Sense Amp as a comparator to perform the digitization using a serial flash, implemented in memory. Flash ADCs require an ordered set of reference voltages to compare against the input to be digitized. Recognizing that the MAC value is non-uniformly distributed and is application specific we propose an algorithm to generate the reference voltages tailored to the MAC distribution function. Further, we show that the reference voltage can be generated in much the same way as the MAC voltage is generated along a column, in-memory. We provide an algorithm to populate the bit-cells of the reference column to generate the appropriate reference voltage. Experiments on the MNIST,SVHN and CIFAR-10 data sets show that the proposed technique results in a worst case accuracy reduction of 0.8% compared to the Double-Precision evaluation.