GDOT: A graphene-based nanofunction for dot-product computation

Abstract

Though much excitement surrounds two-dimensional (2D) beyond CMOS fabrics like graphene and MoS2, most efforts have focused on individual devices, with few high-level implementations. Here we present the first graphene-based dot-product nanofunction (GDOT) using a mixed-signal architecture. Dot product kernels are essential for emerging image processing and neuromorphic computing applications, where energy efficiency is prioritized. SPICE simulations of GDOT implementing a Gaussian blur show up to ~104 greater signal-to-noise ratio (SNR) over CMOS based implementations - a direct result of higher graphene mobility in a circuit tolerant to low on/off ratios. Energy consumption is nearly equivalent, implying the GDOT can operate faster at higher SNR than CMOS counter-parts while preserving energy benefits over digital implementations. We implement a prototype 2-input GDOT on a wafer-scale 4" process, with measured results confirming dot-product operation and lower than expected computation error.