Unlike most available RRAMs and memristors which only offer binary operation the PSU team have demonstrated a non-volatile graphene-based resistive memory device which is capable of achieving in excess of 16 conductance states. One commercial startup, Weebit Nano Ltd., has filed a patent for multi-level cell operation of its silicon-oxide based ReRAM (see Weebit, Leti file multi-level ReRAM programming patent).
With binary memories circuit designers must either use multiple bits to replicate a high-resolution analog signal or make use of more complex networks that will tolerate binary operation but the rounding of trained weights introduces error which can limit inference accuracy.
Where conductance is highly programmable it can be used for weighting signals while keeping networks as simple as possible.
"We are creating artificial neural networks that seek to emulate the energy and area efficiencies of the brain," explained Thomas Shranghamer, a doctoral student and first-named author on a paper recently published in Nature Communications.
The paper also shows that the graphene memory shows the necessary retention and programming endurance to enable weight assignment based on k-means clustering, which offers greater computing accuracy when compared with uniform weight quantization for vector matrix multiplication.
The authors used a graphene field effect transistor (GFET) as the non-volatile memory. These were manufactured by chemical vapor deposition (CVD) of graphene onto a 50 nm alumina (Al2O3) substrate, which acts as a back-gate oxide, on a stacked Pt/TiN/p-typeSi, back-gate electrode. Each GFET had a channel length (L) of 1-micron and a channel width (W) 0.5-micron.
The GFET was programmed into 2, 4, 8, and 16 conductance states, respectively, through the use of different write pulse step sizes.
The team claims that ramping up this technology to a commercial scale is feasible. The US Army Research Office supported this work. The team has filed for a patent on this invention.
Related links and articles: