Talk given by Prof. Shahar Kvatinsky
Today’s AI applications demand tremendous computing power. However, existing hardware for AI is hitting a bottleneck in terms of speed and power. This is particularly because of the data movement required due to the separation of computing and memory in the von Neumann architecture and the large energy consumption, access time and cost of existing memory technologies. Existing AI uses devices like GPUs and dedicated hardware like TPUs and edge inference devices like ASICs. However, the brain is able to perform its processing at very low power and is particularly good at perception tasks.
In this talk, Prof. Shahar Kvatinsky explained potential ways to improve hardware for AI by brain inspired neuromorphic computing using emerging memory devices called memristors. Memristors can emulate synaptic functions and can be used to accelerate neural networks. One way is to perform Vector Matrix Multiplication (VMM) and Multiply and Accumulate (MAC) operations using crossbars of memristors. He explained training of memristive neuromorphic systems using backpropagation and stochastic gradient descent, methods for low power neuromorphic computing like low precision AI inference and trainable data converters and discussed some of the security issues and mitigation strategies. He also talked about various emerging memory devices that can be used to build these systems like Magnetic Tunnel Junction (MTJ) based devices and floating gate flash memory called Yflash, built by Tower Semiconductor. He described quantized deep neural networks with MTJ based devices and simpler neural network models like Deep Belief Networks (DBNs) built with Yflash devices.
Prof. Shahar Kvatinsky’s talk highlighted some of the devices and techniques to accelerate and improve the performance of hardware for AI applications.
Written by Rishona Daniels