Fig. 4: Proposal of a spintronic chip for efficient neuromorphic computing.

a Rending of the envisioned spintronic chip for neuromorphic computing. b Simulated operation of the neuromorphic chip in the bias-enhanced tunnel magneto-Seebeck (bTMS) regime for the task of handwritten digit recognition based on the 800 nm magnetic tunnel junctions characterised in Fig. 3f. The envisioned chip consists of 64 × 10 devices with one row corresponding to an entry in the input vector of dimension N₁ = 64 and one column corresponding to an entry in the output vector of dimension N₂ = 10 (one entry for each possible numeral). c This results in a densely connected single-layer neural network. The pixels of one input image are converted to a vector which serves as input encoded in the laser powers P_j (the laser powers are equal across one row). We apply a bias current I_bias;j,ℓ to each magnetic tunnel junction (MTJ) which serves the role of a trainable synaptic weight W_j,ℓ. d The nonlinear response of the MTJs to the laser power and bias currents enables neuromrophic computing with a high expressivity. The axes have been rescaled with \({I}_{{{\rm{bias}}},\max }=3.0{{\rm{mA}}}\), \({P}_{\max }=150{{\rm{mW}}}\), \({V}_{{{\rm{AP}}},\max }=23.35{{\rm{mV}}}\). The resulting Seebeck voltages are added across one column (dark grey) and (e) result in an output voltage. During training, the weights are adjusted so that the largest output voltage indicates the number shown in the image. Here, for an input image of 7, the network correctly predicts the output after training. f Training and test accuracy for the correct identification of images. g Bias currents (synaptic weights) across the chip after training. The higher the current, the more relevant a pixel for the correct classification.