Rain Neuromorphics has completed a digital processor design
that most closely mimics the human brain. The chip contains
10,000 digital neurons and will be manufactured using a 180nm
process. Scaling up and reducing technology norms will help
create a solution for artificial intelligence problems with
extremely low consumption and powerful cognitive functions.
As you know, the human brain is a neural network of approximately
90 billion neurons. The input of each neuron is a set of "hairs"
called dendrites. A potential (voltage pulse) propagates through
them, which makes the neuron react in one way or another. For a
neuron to react, an averaged sum of potentials of all signals from
many dendrites is required, which makes the brain an "analog
computer".
The response of a neuron is also represented by an action potential,
but it spreads along other nerve "hairs" - axons. As a rule, neurons
have one axon (output), which ends in a synapse - the place of
interaction with the dendrites of other neurons, if we talk about
the brain. In a chip, brain chemistry is either difficult or
impossible to replicate. Therefore, imitation will always be
conditional and the Rain Neuromorphics chip is no exception in
this regard. But on the other hand, progress has been made in terms
of the reproduction of dendrites. According to the developers, the
first chips will be able to provide 125 million INT8 parameters for
processing vision, speech, natural language and recommendations,
while consuming less than 50 watts. The company expects samples
to be available in 2024, with silicon ready for commercial supply
in 2025.
