Comments from cea Leti to show advancements that are making on...

Comments from cea Leti to show advancements that are making on AI using WBT reram.

The human brain has a very high computing efficiency for applications like recognising forms or faces. It performs these operations using only 20W. Another example is the bee’s brain, which does very little computing, but is still very efficient compared with integrated circuits. Biological systems have become a source of inspiration for the semiconductor industry.

We tried to mimic the behaviour of the brain, with synapses and neurons, using non-volatile memories and transistors to reach the same order of energy efficiency.

We first launched a neuromorphic chip called Spirit. It was a demonstrator, using quite old technology – 130nm CMOS technology, with integrated non-volatile memories. We had only 10 neurons and 144 synapsis on this demonstrator – and the circuit consumed only 3.6 picojoules per synaptic event. That’s around eight or 10 less than the Intel demonstrator and 10 times less than the IBM demonstrator.

The computational task we had this chip perform was to recognise handwritten digits. We disclosed this chip at an IBM conference in 2019.

Now we are preparing the next generation, which will be a scaled-up version of this approach. We plan to have more than 100,000 neurons on a chip and more than 75 million synapses. This is a promising area, with a potential for huge gains in power efficiency.

Exciting times ahead on all fronts.