Penta - Transistor Integrate & Fire (PTIF) Spiking Neuron with an Ultra-Low Energy Consumption of 0.045 fJ per Spike

• Published in: Conference proceedings : Midwest Symposium on Circuits and Systems pp. 1060 - 1064
• Main Authors: Williams, Shelby, Khalil, Kasem, Bayoumi, Magdy
• Format: Conference Proceeding
• Language: English
• Published: IEEE 11.08.2024
• Subjects: 45nm technology mode; Brain modeling; Computational modeling; Energy consumption; femto.Joules; Integrate & Fire (I&F); MOSFET; Neurons; Neuroscience; Semiconductor device modeling; Spike Neural Network (SNN); subthreshold supply voltage
• ISSN: 1558-3899
• DOI: 10.1109/MWSCAS60917.2024.10658774

The human brain's computational efficiency reigns supreme. There is no known device that even comes close to its low-power, massive parallelism or its data-processing capabilities. The ultimate goal of neuroscience is to mimic the brain. The proposed analog circuit is one such attempt. It utilizes an artificial, silicon-based, neuronal implementation of the widely recognized Integrate & Fire (I&F) spiking model. Its topology exhibits superior structural simplicity, comprising only five equivalently-sized MOSFET transistors. This Penta Transistor Integrate & Fire (PTIF) model uses a positive-feedback design. Energy efficiency is the key distinguishing hallmark of this PTIF circuit, with an energy consumption of only 0.045 femtojoules (fJ) per spike. In fact, to the best of our knowledge, the proposed PTIF implementation has the lowest energy usage per spike than any previously reported Integrate & Fire (I&F) neuron design by one magnitude (1Ox). Linear and quadratic energy-saving performances stem from the exclusion of external capacitors and the utilization of a subthreshold supply voltage, respectively. Additionally, no complex transistor resizing is necessary, allowing for easy transfer to much smaller technology nodes. The aforementioned attributes of the PTIF design make it highly advantageous for ultra-low-power applications, especially well- suited for Edge Computing (EC), Federated Learning (FL), and Internet of Things (IoT).