Energy efficient single flux quantum based neuromorphic computing
Michael L. Schneider, Christine A. Donnelly, Stephen E. Russek, Burm Baek, Matthew R. Pufall, Peter F. Hopkins, William H. Rippard
Many neuromorphic hardware technologies are being explored for their potential to increase the efficiency of computing certain problems, and thus facilitate machine learning with greater energy efficiency and or with more complexity. Among the technologies being developed, single flux quantum based Josephson junctions are a promising choice for their extremely low energy consumption and intrinsic spiking behavior. Recent experimental work has demonstrated nano- textured magnetic Josephson junctions (MJJs) that exhibit tunable spiking behavior with ultra- low training energies in the aJ regime. MJJ devices integrated with standard single flux quantum neural systems form a new class of neuromorphic technologies that have spiking energies between 10-18 J and 10-21 J, operation frequencies up to 100 GHz, and nanoscale plasticity. Here, we present the design of neural cells utilizing MJJs that form the basic elements in multilayer perceptron and convolutional networks. We present SPICE models, using experimentally derived Verilog A models for MJJs, to assess the performance of these cells in simple neural network structures. Modeling results indicate that the tunable Josephson critical current IC can function as a weight in a neural network. Using SPICE we model a fully connected two layer network with 9 inputs and 3 outputs.
2017 IEEE International Conference on Rebooting Computing (ICRC)