Design-technology co-optimization for OxRRAM-based synaptic processing unit

TLDR

A primary bottleneck for using a single non‑volatile memory as a synaptic unit is the need for a high number of distinct weight levels. This work presents a design‑technology trade‑off analysis for implementing a fully connected neural network with OxRRAM cells, introduces a mixed‑radix encoding for multi‑device synaptic units that attains 94 % accuracy despite device variability, and is the first to examine single versus multi‑device weight trade‑offs using silicon data. The authors evaluate the trade‑offs on a 1 Mb OxRRAM array, employing mixed‑radix encoding to map synaptic weights across multiple devices. The mixed‑radix scheme achieves 94 % classification accuracy and the neuromorphic algorithm successfully mitigates high device variability.

Abstract

In this paper, we present a design-technology tradeoff analysis to implement a fully connected neural network using non-volatile OxRRAM cells. The requirement of a high number of distinct levels in synaptic weight has been established as a primary bottleneck for using a single NVM as a synaptic unit. We propose a mixed-radix encoding system for a multi-device synaptic unit achieving high classification accuracy (94%) including device variability. To our knowledge, this is the first paper to discuss the tradeoff between single and multi-device synaptic weight in terms of design and technology using silicon data. We have demonstrated that high level of variability can be handled by the neuromorphic algorithm. The results presented in the paper has been obtained from 1Mb array.