ASIC Implementation of Efficient 512-Neuron 256K-Synapses Digital Neuromorphic Processor with On-Chip Encoding in 22nmFDX

There is a rising demand for AI workloads running on the edge that are increasingly complex, calling for more efficient computing platforms. Neuromorphic hardware is a class of electronic circuits and devices that aim to mimic a biological brain’s functionality and computational efficiency. In this work, we propose a 512 neuron 256K synapses LIF-based digital neuromorphic processor. Each neurosynaptic core can be configured to act as either an encoding neuron or as a processing neuron, chosen to be Leaky-Integrate-and-Fire (LIF). We have also implemented a 1-1 software code for the processor, in which Pytorch and SNNtorch are used to build the spiking neural network, and surrogate gradients are used for training. An image classification task is used to evaluate the performance of the neuromorphic processor using the MNIST dataset, achieving an accuracy of 96.12% for 4-bit quantized synaptic weights. Implemented using GF22nmFDX technology, the processor is 0.72mm
²
in area, of which synaptic memories occupy 68% of the area, and achieves an average power efficiency of 14.8 nJ/SOP. This large area is justified by the configuration flexibility of the processor, where each neuron can be connected to any other. This allows the mapping of different and more sophisticated tasks.