IBM has made two prototype neural net ICs each with 256 ‘neuron nodes’ which, in one chip connect to 262,144 synapse-type memories and, in the other chip, connect to 65k synapses.
They were fabricated using a 45nm SOI-CMOS.
Synapse aims to replicate the computing power of the human brain within the size and power consumption restraints of the human brain.
IBM calls its two prototype chips ‘neurosynaptic cores’ aimed at ‘cognitive computing’.
The ideas behind neural nets became formalised in the 1940s – Alan Turing worked on the theory of how they could be implemented – but they have never achieved the computational horse-power to perform anything but rather limited tasks.
A host of companies have been started to bring the technology to market.
In 1986, VLSI guru Carver Mead and microprocessor pioneer Federico Faggin formed a company called Synaptics to pursue neural nets, and had some success in PC touchpads and other applications.
Axeon in the UK used neural network technology developed at AberdeenUniversity to control car engines.
Many, many more have tried to apply the technology, and universities have never wavered in their efforts to develop it.
However neural networks have never realised their potential of getting anywhere near to what the human brain does in tasks like pattern recognition, lateral thinking, self-learning and other brain-like functions.