A team from IIT Bombay developed and demonstrated a spiking neural network chip. The chip consists of an input layer of 20 spiking neurons transferring excitation to a reservoir of 36 neurons. The reservoir is randomly and recurrently connected through static binary (0/1) synapses. This is a demonstration of the reservoir of a liquid state machine.
The chip supported the application of spoken word recognition. This application requires that information spread in time (a time series) be stored in memory and then sent to the processor in chunks of a few seconds at a time. This need for continuous data storage and transfer between memory and the processor takes a lot of power. The spiking neurons can process in real time without storing and recalling the data from memory.
The neurons in the chip depend on very stable and low current sources, known as quantum tunneling. Such tunneling occurs in transistors in the off-state. The tunneling currents are about 1000x lower than typical on-state currents, enabling a very low power operation. Further, the neurons operate at 100 kHz without bulky capacitors, enabling a compact design. This is close to the ~1 kHz data rate of human speech.
The chip is fabricated in 45nm RF SOI technology from the GlobalFoundries University Program and funded in part through MeitY and DST projects.
Demo Type: IC Design, Device
Faculty: Udayan Ganguly, Maryam Shojaei Baghini
Institute: IIT Bombay
Govt. Funding Agency: MeitY, DST
Industry Engagement: GlobalFoundries
Year of Publication: 2022
Reference:
- Ajay Singh[1] , Vivek Saraswat, Maryam Shojaei Baghini, Udayan Ganguly “Quantum Tunneling based Ultra-compact and Energy Efficient Spiking Neuron enables Hardware SNN” IEEE Trans. Circuits and Systems-1, 2022 (Preprint, Final )
Group Website:
Udayan Ganguly, Maryam Shojaei Baghini
Popular Article: IEEE Spectrum
Video: YouTube Link
Keywords:
Spiking Neural Network
Silicon on Insulator (SOI) Technology
Quantum Tunneling Neurons
Liquid State Machine,
Spoken Word Recognition