Intel announced near the end of 2021 that their second-gen Loihi chip is under development. Although the company believes that large-scale neuromorphic artificial intelligence is still several years away, they’ve been investing in the software, hardware, and development for it during the past four years.
Neuromorphic computing works to replicate the brain’s circuitry. Neurons communicate by sending spiked signals across synapses to each other. The brain interprets the inputs sent by the senses, delivering output to the muscles.
At the moment, a 100-billion parameter AI requires a 1,000-GPU cluster and several months to train. It works out to over a quadrillion neural network computations per second for an entire day. That means it takes more than a year to generate reasonable text.
It’s $4 million to train a network once. That’s why Loihi 2 is an exciting concept.
Loihi 2 Fixes the Limitations of the Original
The issue with Intel’s first attempt with the Loihi technology was that the spike signals were not programmable. They had no context or value range. The second-gen chip solves both those limitations while offering faster circuits, more neurons, and additional link bandwidth.
That means the AI scalability is much faster.
It’s been built on the pre-production version of the Intel 4 process. The simplified design rules and EUV technology have helped the innovation processes happen faster, creating the potential for commercialization.
That’s why Intel thinks it is several years away from having an extensive framework for Loihi 2. Everything needs to be rebuilt from scratch.
Once those foundations are in place, the scaled-up systems will work faster, use data more efficiently, and help us all to be more productive. Innovations like these in the technology world seem to be escalating.