From: Eugene Leitl (Eugene.Leitl@lrz.uni-muenchen.de)
Date: Thu Sep 06 2001 - 14:41:53 MDT
On Thu, 6 Sep 2001, Charles Hixson wrote:
> I tend to believe that eventually an AI would need to swallow the OS
> of the system on which it was implemented. That's why Linux is a
Sure, but you can't implement AI on current hardware, for the same reason
you can't implement a naturally intelligent AI on a paper tape/trained
termite Turing machine. It's too s l o o w.
> better choice. Or *BSD. Or something else, nearly anything else,
> that had source code available AND mutable AND redistributable. It
Beowulfs are useful, and scale to several 1e3 nodes, but beyond that
footprint, power, air conditioning and hardware failure rate begin to turn
it sour. Here you have to go custom (and very elephantine, since no
longer COTS), or give up. Given current budgets, one should stick with
COTS.
> would require a much higher level AI to swallow a binary OS, and the
> development would be much more difficult in the absence of any OS. I
The OS is largely a nonissue, you need something reliable and good QoS in
regards to the signalling layer. The message passing library itself can
(and should) be user space. If we want to talk about the best wood to use
for an abacus. But, I'd rather not to use an abacus to render a movie.
> really doubt that any general purpose AI will be developed on special
> purpose hardware. After it has been developed, it may decide that for
> some functions specialized hardware is superior, despite the higher
> cost. Or it might not (outside of sensory modules, etc).
You're arguing from the position of the year 2001. Currently, you're
right, the best ROI is COTS, and will continue to be that for quite a long
time. As much as another decade, maybe. However, the difference between
Moore's law and real world performance is starting to gap perceptibly, and
there's some trend towards diversification, or rather, specialization. So,
we will have to have special-purpose hardware, ere the noughties will have
grown very old.
> Since at the most recent Linux World conf. I saw a plug-in
> board with a plug in 25GB disk card (combined price ~ $700), and
Hard disk? What do you need hard disk for? It's much too slow for anything
but checkpointing. Hard disks will die, soon.
> the cpu was a 2 GHz chip and the boards could be connected with
Nice, but today we have 20 GHz FPGAs, to be boosted to 50 GHz this year,
and 100 GHz next year.
> a ethernet, and run as a Beowulf complex. You could, in
> principle, fit as many into your computer as you had slots for
> (and your power supply would support). That's NOW. That's on
> the market today. I feel relatively certain that when the
> software is ready, hardware won't be the problem.
Beowulfs are nice, but the hardware price becomes irrelevant at about 10 k
nodes, where other factors start to dominate.
> Of course, it would be a bit expensive... at today's prices. To take
> a wild fling: Say you have one tower box with four free slots, that
Slots: bus. Shared. Not crossbar. Bad.
> you fill this way. That's $2000 + 4 * $700 ~ $5,000. Now this isn't
> really enough, so say you connect five of these together. $25,000.
> That's a good starter system. And the OS for this would be Linux. I
> couldn't put a good price estimate on what it would cost to do this
> with Windows, but since that's 25 separate cpu's I think it would
Windows is a truly lousy OS for a Beowulf, for multiple reasons.
> probably about double the cost. Or more. And for a better system one
> would want to cluster five of these groups together... etc.
Sure, but I thought we were talking AI, not Beowulfery.
> Well, today's too early to start buying the hardware. Moore's law
Not if you want to get something done, today.
> says do as much as is feasible before buying the hardware. If I had
Moore's law is about integration density, not performance.
> gotten far enough, I'd spring for the basic Tower ($5,000). That
> would let me start working on the main thread dependency issues, etc.
Thread? ???
> By the time I was ready to proceed, hareware will probably have
> changed enough that it would be better to start over on new hardware.
> But Linux ports easily, so that's no problem. On the 5X5 the software
Yes, there is, if you want to port your code to a 1e6 node machine.
Because Linux won't run on that, for obvious reasons.
> would probably become sophisticated enough to swallow the OS, so one
> would need to be sufficiently convincing to cause it to hold off until
> one could add at least a second 5X5 module. Then it should understand
> that 5X5's could be added or removed, and design itself accordingly.
You're very fond of numbers. I used to do that, too.
> If I project this year's trends forwards (cpu speeds doubled,
> prices fell a tiny bit) then in a couple of years one should be
CPU speed growth over time is no longer a linear log plot. And memory
bandwidth is most assuredly not that. So screw Moore.
> able to get a 5X5 that was 4 times as powerful for around
> $15,000. (Actually, it would do better than that. Multiple
> CPUs don't scale nicely in capacity. It depends on the problem
SMP does lousy on memory starved code. And AI is most assuredly that.
Properly done AI is embararssingly parallel, so scaling up cheap nodes
with good interconnect is your best bet. Currently.
> mix, but generally you're doing quite well if you get a 50%
> bonus by adding the second one. So you gain quite a lot more by
> using faster CPUs than by using more of them.)
Your memory bandwidth is still screwed, so fuck that.
> And do remember that an AI won't need to be as busy at the
> autonomic tasks as we are. I've seen an estimate that 80% of
Can't parse this. Autonomic tasks?
> our brain is devoted to tasks like maintaining muscle tension,
> processing food, monitoring sugar levels, etc. I don't know how
You assume the AI won't have sensomotorics???
> reliable the estimate was, but current hardware doesn't need
> much of that kind of effort. So a much less computational
The estimate is screwed, most is burnt to process vision. At least, if
we're talking about hardware between our ears.
> intensive AI would be needed to achieve the same results. And there's
> also the question of how efficient our thought processes are. I don't
Judged from past estimate history (aka perpetuation of debacles), you must
assume they're about as efficient as it gets, plus some, and then twice
that.
> know of any way to answer this, but there may be a lot of room for
> improvement. In fact, we may have already designed a bunch of
Improvement for what? You don't have to track single molecules in the
brain, that is more or less certain. But, you certainly have to account
for mentation-relevant processes occuring in an instance of a natural
nanocomputer, and this is anything but paltry stuff. You better believe
it.
As I said, the space between your ears is very busy, and on a fraction of
the energy footprint of your desktop box. Do not underestimate this stuff.
> improvements that just don't happen to work very easily with a neural
> net. (Symbolic logic, e.g.) We can "understand" them, but they sure
> aren't native to us, so we "run them under emulation". That's always
> slow!
What makes you think AI is pure symbolic logic? If I need that, I can get
Mathematica, or Maple, or MuPad. It just kills me, but this is not natural
AI.
> So I don't see hardware as being the problem (though if it takes
> the 125 node cluster to do the job, affording it might be). But
Hardware is not a problem, a 125 node Beowulf will do!
> notice that this scenario assumes that one is starting for here
> and now. And it requires an OS be present that can BE swallowed
If you want to start here and now, Beowulf is the ticket. But, you're not
going to get any AI that way.
> by the AI, without it being dismembered by lawyers. And it
Lawyers dismembering AI? Strange.
> requires that the OS be copyable, without the AI being
> dismembered by lawyers. Etc. If you expect that the eventual
Linux is free, and it performs nicely enough, but what has this to do with
a naturally inteligent system?
> software will be sold, then one of the BSD variations should
> probably be your starting platform, but that feels to me like
> slavery. It's one thing to ask it to earn its keep, but it's
> another to sell it. This is likely an irrevelent objection, as
> I don't really see it being constrained very long once full
> sapience is achieved. But it's a felt objection. So I would
> prefer a Linux base. (Besides, I'm more familiar with it.)
I'm feeling like it's 1960, or something.
-- Eugen* Leitl leitl
______________________________________________________________
ICBMTO : N48 10'07'' E011 33'53'' http://www.lrz.de/~ui22204
57F9CFD3: ED90 0433 EB74 E4A9 537F CFF5 86E7 629B 57F9 CFD3
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:10:24 MST