Who needs a CPU? Phase change memory acts as an analog computer (2017) | Hacker News Reader

Who needs a CPU? Phase change memory acts as an analog computer (2017)

arstechnica.com

60 pointsjasondavies11 comments

I am doing an internship in the lab where the cited paper was produced right now :-D

If anyone has questions, I'll do my best to answer or relay(maybe tomorrow though as it is late here). There also was a related discussion about 2 months ago:

https://news.ycombinator.com/item?id=16894818

Can you summarize the nature of the computations? I would guess this is stochastic integration. Where is the bandwidth passing information between computed results?

Could I get a copy of the pre-print?

Nevermind, sorry. The link to Nature wasn't functioning when this was first posted.

Is analog computing reproducible? Digital computing is kind of founded on the idea that running the same program twice will give the same results (i.e. determinism), but it seems to me that analog computing might simply fail to make that guarantee. Is that correct? And if so, have the implications and workarounds been investigated at all?

I expect within bounds... this already exists in floating point operations. Look up distributed games programming, x86, PPC etc for some meaty details, but the gist of it is to never trust the exact value, rather look for values within a suitable range.

Interesting take though - with analog computing, this problem would be everywhere rather than in certain special cases. There probably is some funky math we can use to describe and talk about on HN in the year 2029?

That it would be everywhere was exactly my point; I started listing all the sources of unpredictability in my comment but then I figured that wasn't necessary so I didn't go on that tangent. But yes, threading, floating point, RNGs, cache timing, etc. are all sources of randomness we deal with, but a computer where pretty much /any/ computation has to deal with randomness is a whole new ball game so far as I can tell.

This reminds of me Chris Domas' compiler that always outputs the same text segment:

https://github.com/xoreaxeaxeax/reductio

Hacking possibilities increase if even simple devices turn out to be Turing Complete. "Help, my flashlight emptied my bank account!"

> if even simple devices turn out to be Turing Complete

It's not hard to find "simple" devices that are turing complete, 2d cellular automata (with the correct rule) are turring complete for example and they are incredibly simple.

This is related to a notion Wolfram has been exploring and trying to explain for some time, which is that contrary to our previous beliefs (or most people's current beliefs especially the engineers perspective), complex devices are not needed to produce complex behaviour, in nature it is very common for simple systems to produce complex behaviour, including the ability of simple systems to compute... to the point that often these systems look like random noise might have been introduced, but in fact they are completely deterministic (think systems exploited by PRNGs).

"in nature it is very common for simple systems to produce complex behaviour, including the ability of simple systems to compute"

I'm reminded of another recent HN thread about the "intelligence" of single living cells. Whether or not you think they are actually "intelligent", they are clearly capable of complex behavior without being governed by rigid parameters and yet are obviously molecular machines that seem tantalizingly close to being analyzable by humans.

Reply on news.ycombinator.com