The shared nothing message passing architecture is a good fit for browser web workers. I recently tried training a mini-transformer in the browser and got it's training time from 3 minutes to 15 seconds by spawning web workers to do parallel training and then just fusing the independently learn parameters . There is a whole multi-core paradigm in the browser we don't often exploit that does saturate my computer and makes the fans turn on.

I think that multi-core hardware and the extra needed software is overkill for this use-case (share-nothing). Could we expunge all of the implicit caching hardware and software, save silicon and repurpose the saved silicon to put more share-nothing cores on a chip?