2020 BRN Discussion, page-3876

I came across this new test chip from Imec (innovation hub in Belgium) that can compete with current AI accelerators, on just a 4 mm2 die. It performs calculations in analog circuitry and the chip can perform 2.900TOPS/W. They expect to be able to perform 10.000 TOPS/W. It's on 22 nm process node, a much finer process node than CEA-Leti at 130 nm. I read somwhere that if you want to go to 7 nm and below, you can't have analog circuitry.
https://www.globenewswire.com/news-release/2020/07/08/2059469/0/en/Imec-and-GLOBALFOUNDRIES-Announce-Breakthrough-in-AI-Chip-Bringing-Deep-Neural-Network-Calculations-to-IoT-Edge-Devices.html

So I'm a little in doubt to which degree this would compete with Akida, but I suspect it's not going to be for some time at least.

First of all, it doesn't seem to be production ready and probably will not be for some time.

Second, I suspect that it's harder for other companies to integrate analog circuitry into their chips, even impossible if they are going down to 7 nm or below.

Third, it's hard to estimate the performance of the chip, as the mentioned results are in TOPS/W and not TOPS.

Fourth, it's still only acceleration of software emulated neural networks.

To me it still makes sense that the winner of the AI race will be digital and not accelerated software.