Analog is the future of PC's and AI ??

[Eban]

Might be long to watch but very interesting!

[duskw4lker]

so if I understood it right, the key is the weight assigned through 'training' the 'analog neurons' by substracting errors where each 'analog neuron' that outweighs or gets outweighed by other layers of neurons value in a particular case of defining which criteria/output/answer is delivered/applied through a matrix calculation that ultimately represents a quorum that reached an agreement??... what?? I'm not trying to be a smartass, it was really complex to try to comprehend all that info, and I'm glad that I don't have to solve those kind of problems 😅

 

however it may be, it's great to see how technology goes back to the basics (repurposing flash cells currently used for digital storage to analog calculations), and greatly improve those concepts in the pursuit of tackling a particular problem. Different tools for different problems, awesome 😄

[Ishayin]

Thanks for posting – that is interesting indeed and a lot of different aspects to consider. I feel like the whole "exact vs inexact" dichotomy here is a bit well, inexact. The question should surely rather be "how exact". From a physics perspective, I assume the inexactness they are speaking to is associated with the noise levels inherent in the electronic signals and read-outs. So presumably they could do things like reduce the temperature to reduce the Johnson/thermal noise, or increase currents/voltages, or add duplicated parallel processing with balanced connections. Though all of these options would increase power requirements. Anyway, I figure there must be some kind of trade-off that can be had with different designs with different compromises suited to different applications...

 

Anyone here a computer scientist who could speak to what the fundamental application limitations might be? From the very simplified story presented there you get the idea that the range of possible applications should be much broader than just AI, but maybe that is misleading. Surely the output of GPUs for example does not need to be exact, so can we utilise an analog chip for at least the final rendering part of the computational chain? If it's not possible to develop general purpose analog chips, then they won't replace CPUs, but perhaps more promising as additional complementary chips that can be assigned particular tasks?

 

Anyway, always nice to see the development of much more efficient approaches, and people generally thinking outside the box. ☺️

[aldarxt]

Very Interesting even groundbreaking kinda like when man added the wheel to the box, he no longer needed to carry the box, he got in and went

[Eban]

Even those morons at micro$lop see potential for analog computing.

TLDR: calling it AIM analog Iterative Machine. Sidesteps mores law using optics and speed of light and 100x faster than digital (they say).

 

https://www.techspot.com/news/99228-microsoft-has-analog-optical-computer-doesnt-use-transistors.html

[Sandy Bridge]

Wired had an article on this topic in one of the recent issues (I get the paper version, nice to read something that isn't on a screen with breakfast in the morning).  It's been a month or two and I'd say my understanding is fuzzy because of that, but that isn't the only reason my understanding is fuzzy, it's also a challenging concept.

 

What I do recall is that making fast analog CPUs general-purpose programmable is currently not easy.  The analogy they used was that we're basically in the 1950s still, there's been a grad student who did a partial Python port to a boutique fast analog CPU that a Columbia professor has made about 20 of, but we're still waiting for our FORTRAN-on-a-common-chip that can really open the floodgates.

 

They mentioned at least one potential use case that isn't neural nets.  I think it was something physics related, which would make sense; analog computers were responsible for the rangekeepers on battleships such as the Iowa class (1943 - 1991).  And maybe narrowing down the range of possible results, for final refinement by a digital computer, and at much lower overall energy expenditure than if the calculations were done purely digitally.

 

Based on my admittedly likely-inaccurate understanding, I think part of what Ishayin said may be right - even if these chips don't become wholly general-purpose like current CPUs, they may be very useful in certain domains.  A battleship rangekeeper didn't need a Ryzen 7950X with 16 high-performance general-purpose cores that could also run Crysis, it "just" needed an analog computer that could calculate where to point the guns to hit the target ("just" in quotes because that's not an easy physics problem).