Principal Investigator Jeehwan Kim
Although several types of architectures combining memories and transistors have been used to demonstrate artificial synaptic arrays, they usually present limited scalability and high-power consumption. Analog-switching devices may overcome these limitations, yet the typical switching process they rely on, formation of filaments in an amorphous medium, is not easily controlled and hence hampers the spatial and temporal reproducibility of the performance.
Here we demonstrate single-crystalline SiGe epiRAM with minimal spatial/temporal variations with long retention/great endurance, and high analog current on/off ratio with tunable linearity in conductance update, thus justifying epiRAM’s suitability for transistor-free neuromorphic computing arrays. This is achieved through one-dimensional confinement of conductive Ag filaments into dislocations in SiGe and enhanced ion transport in the confined paths via defect selective etch to open up the dislocation pipes. In SiGe epiRAM, the threading dislocation density can be maximized by increasing Ge contents in SiGe or controlling degree of relaxation23, and we discovered that 60 nm-thick Si0.9Ge0.1 epiRAM contains enough dislocations to switch at tens of nanometer scale devices. The simulation-based on all thosecharacteristicsofepiRAMshows95.1%accurate supervised learning with the MNIST handwritten recognition dataset. Thus, this is an important step towards developing large-scale and fully-functioning neuromorphic-hardware.