Materials challenges for non-silicon matrix multipliers and neuromorphic computing*

Workload driven changes in computing have had two major hardware consequences: (i), the emergence of new types of non-volatile memory; and (ii), dedicated hardware for executing neural network algorithms. Items (i) and (ii) are related, and new types of nanoscale memory and selector switch elements based upon phase change, magnetic, electrochemical, and insulator-to-metal (IMT) transition phenomena are being investigated as cross-bar memories, in-memory computing for matrix multipliers, and neuromorphic circuits that require devices mimicing synaptic and artificial neuron functions. I will first define—from the materials science perspective—performance characteristics of synaptic and neuronal devices that will need to be met to satisfy the needs for neuromorphic computing and the building of matrix multipliers. Using simple models based upon diffusion, surface tension and drift in conducting filament devices, I will examine the differences between selector switches and non-volatile memories and highlight the materials characteristics desirable for these two types of devices, and the limits on their variability. I will also describe our results in successfully building low voltage (<500 mV) artificial neurons using IMT materials, and low voltage conductive filament memory devices that use ultraporous matrices based upon a new synthesis technique called sequential-infiltration-system (SIS) [1].

[1] Q. Peng, Y. C. Tseng, S. B. Darling, J. W. Elam, ACS Nano, 2011, 5 (6), pp 4600–4606.