Powering AI at the edge: A robust, memristor-based binarized neural network with near-memory computing and miniaturized solar cell

Abstract

Memristor-based neural networks provide an exceptional energy-efficient platform for artificial intelligence (AI), presenting the possibility of self-powered operation when paired with energy harvesters. However, most memristor-based networks rely on analog in-memory computing, necessitating a stable and precise power supply, which is incompatible with the inherently unstable and unreliable energy harvesters. In this work, we fabricated a robust binarized neural network comprising 32,768 memristors, powered by a miniature wide-bandgap solar cell optimized for edge applications. Our circuit employs a resilient digital near-memory computing approach, featuring complementarily programmed memristors and logic-in-sense-amplifier. This design eliminates the need for compensation or calibration, operating effectively under diverse conditions. Under high illumination, the circuit achieves inference performance comparable to that of a lab bench power supply. In low illumination scenarios, it remains functional with slightly reduced accuracy, seamlessly transitioning to an approximate computing mode. Through image classification neural network simulations, we demonstrate that misclassified images under low illumination are primarily difficult-to-classify cases. Our approach lays the groundwork for self-powered AI and the creation of intelligent sensors for various applications in health, safety, and environment monitoring.

Fig. 1. Overview of the fabricated memristor-based binarized neural network.

a Optical microscopy image of the fabricated die, showing four memory modules and their associated digital circuitry and power management unit. b Detail on one of the memory modules. c Cross-sectional scanning electron micrograph of a hybrid CMOS/memristor circuit, showing a memristor between metal levels four and five. d Schematic of a memory module. For each operation mode, biasing conditions for WL, BL, and SL are given with respect to the power domain (VDDC, VDDR) and VDD. e Schematic of the level shifter, used in d for shifting digital voltage input to medium voltages needed during programming operations or nominal voltage during reading operations of the memristors. f Schematic of the differential pre-charge sense amplifier PCSA, used to read the binary memristor states, with embedded XNOR function, to compose a XPCSA: it computes an XNOR operation between input activation X and weight (memristor value) during bit-cell sensing.

Fig. 2. Measurements of the memristor-based binarized neural network, employing a lab-bench power supply.

a Sample measurement of the output of the integrated circuit, compared with a delay-less register-transfer level (RTL) simulation. b Photograph of the printed circuit board used for the experiments. c Measurement of the energy consumption to perform a whole-chip inference, for various operating frequencies and supply voltages. d Pie chart comparing the different sources of energy consumption in the system, obtained using simulations (see Methods).

Fig. 3. Accuracy of the memristor-based binarized neural network.

a Measured schmoo plot, presenting mean accuracy of the output neuron activations, for different operation frequency and supply voltage. They were obtained using patterns of weights and inputs chosen to cover all possible neuron preactivations (see Methods). NF means non-functional. b Measurements of 64 neurons with preactivations –5, –1, 0, 1, and 5, at 33 and 66 MHz with a power supply of 0.9 volts. Errors are marked in red. Mean accuracy of the output neuron activations, as a function of neuron preactivation Δ and supply voltage, measured at (c) 33 and (d) 66 MHz (see Methods).

Fig. 4. Measurements of the binarized neural network powered by a miniature solar cell.

a Schematic view of the AlGaAs/GaInP heterostructure solar cell. b Photograph of the solar cell, and its measured current-voltage characteristics under one-sun AM1.5 illumination provided by a standardized solar simulator (see Methods). c Current-voltage characteristics of the solar cell for various illuminations provided by the halogen lamp (see Methods). d Photograph of the experimental setup where the fabricated binarized neural network is powered by the solar cell illuminated by the halogen lamp. e Mean measured accuracy of the output neuron activations, with the binarized neural network powered by the solar cell, as a function of neuron preactivation Δ and solar cell illumination.

Fig. 5. Neural-network-level investigations.

a, b Illustration of our method for mapping arbitrarily-shaped binarized neural networks to 64 × 128 memristor arrays. The detailed methodology is presented in Supplementary Note 4. c t-distributed stochastic neighbor embedding (t-SNE) representation of the MNIST test dataset. The datapoints incorrectly classified under 0.8 suns (left) and 0.08 suns (right) equivalent illumination, but which would be correctly classified under 8 suns, are marked in black. These results are obtained using a binarized fully-connected neural network (see Methods). d Same graphs as c, obtained for the CIFAR-10 dataset and using a convolutional neural network (see Methods).