Crunching data faster and faster is what computer industry has been doing over the last 60 years, tracking with amazing precision Moore’s law. Since 2014/2015 Moore’s Law has reached its limit, first in the economics (the price per transistor is no longer decreasing) and then in density (we can no longer double the number of transistors in a given area every 18 months). However, the number crunching performance has kept increasing (slowing down a bit…) and this is thanks to new ways of packaging transistors in the chip, new architectures. We have been moving from planar arrangement of transistors to 3D, we have multiplied the number of “cores” (processing units) within a chip and placed memory inside the chip to avoid wasting time in transferring data from storage chip to processing chips and so on.
Another way industry is following to increase chips performance is the design of special chips focussing on specific activities. One example is neuromorphic chips, chips that are copying the interconnection structure of the brain neuronal circuits.
This is the case of the just announce Loihi 2, a neuromorphic chip by Intel, manufactured with a new, advanced, etching process, Intel 4, that will be applied to all Intel chip production starting in 2023. This chip, produced is small quantity is a sort of prototype to perfect this new process.
According to Intel announcement, the Loihi 2 is 10 times faster than the previous version and consume less power. It embeds some 1 million neurones like components that can be wired (via software? on demand, to better fit the task at hand. This is probably the most interesting part, what separates a neuromorphic chip from its siblings: the possibility to decide after the chip has been manufactured, the internal wiring.
This is what happens in our brain that through experience changes its wiring to adapt (minimise the energy budget) to a specific activity. Researchers have demonstrated that the brain of a professional basketball players needs to work only a fraction (hence use a fraction of energy) of an amateur that is struggling to send the ball through the hoops. This is due to the rewiring, occurring after long training, that makes processing of sensorial inputs (vision, muscle coordination) much more straightforward.
Likewise for these neuromorphic chips. You can program them to perform specific activities, like recognising images for an autonomous car. The difficult part (once the chip is available) is in this programming of the chip hardware (once this is done you will need to write the program that will run on the chip…). To simplify the chip programming Intel has created an open source software framework called Lava,
With this software it is possible to set up from 100 to 10,000 connections among any two of the 1 million neurones like components it contains.
Notice the small size of the chip: this means that its cost will be low and hence it can be embedded in many devices at the edge providing local intelligence, rather than tapping on Cloud intelligence.