Quantized Domain-Wall Magnetic Tunnel Junction (DW-MTJ) Neural Networks Optimized for Rapid, Energy Efficient Edge Inference

Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. The increasing complexity of artificial intelligence (AI) models implies a need for new systems that efficiently process deep neural networks (DNNs) [1]. DNNs can efficiently be implemented using crossbars of dense nonvolatile memories (NVMs) [2]. While most DNN NVM implementations focus on ReRAM, Phase-change memory (PCM), and binary magnetic tunnel junctions, domain-wall magnetic tunnel junctions (DW-MTJs) are naturally analog and offer spin-transfer torque and spin-orbit torque operation for logic [3,4] and neuromorphic [5] operations. DW-MTJ devices perform online learning; e.g. interacting neuronal DW-MTJ devices implement winner-take-all computations in shallow learning systems [6,7] . Efficient multi-layer perceptron learning with linear and symmetric notched domain wall synapses has also been demonstrated [8].