From: Robert J. Bradbury (bradbury@www.aeiveos.com)
Date: Mon Dec 06 1999 - 11:05:43 MST
On Mon, 6 Dec 1999, John Clark wrote:
> however today IBM announced they will spend $100,000,000 make a machine of
> that power in just 5 years. This will be called "Blue Gene" and will be
> specially designed to do one thing, not play chess but solve the awesomely
> complex protein folding problem.
Waaallll, just stomp my face right into the mud...
I just posted a note yesterday to the EI List commenting on the discussion
of this in the "Blitzing Bits" (pg 57) article in the Scientific American Extreme
Engineering special issue with my interpretation that IBM had not made
a commitment to this --
"This is a cheap way to build supercomputers if you need them for just
one kind of problem, says "Mark Snir, manager of scaleable parallel systems
at IBM. 'We looked here at what it would take to build a multipetaflops
machine customised to the problem of protein folding," he says. "We could
do it for a few million dollars--much, much less than a general-purpose
machine.'"
and here I find that the next day, they are making a commitment
(presumably after the marketing guys and managers had inflated
the cost a bit).
> IBM thinks there is lots of money to be made in analyzing the huge
> amount of data that will come from the Human Genome Project,
They got that right.
> I think it could also be used to design nano machines.
Phooey... You could design nanomachines now. After all Eric & Ralph
have done specific parts and NASA Ames has simulated actually running
them.
The protein folding problem is *much* more difficult than nanoscale
design because of the large numbers of degrees of freedom you have
in protein folding. For most nanoscale designs I can envision
(e.g. scaling down macromachines or MEMS machines to nanoscale)
you absolutely do not want that. The design space humans work in
has many fewer degress of freedom because our minds can't deal with
anything too complex. Nature on the other hand wants many degrees
of freedom because they are required to successfully explore the phase
space. [In nature most designs don't work, so you have to try a lot
of them to find the ones that do work.]
Now the machine will be useful for two things (a) really good drug
design, because you will be able to model protein flexing much more
effectively and develop drugs that bind much better, or drugs that bind
well to one specific receptor and not 4 other versions of a receptor.
And (b) the de novo design of enzymes, in the event that hard/dry
nanoassembly proves very difficult and we have to go the biotech
route to pseudo-hard nanotech.
>
> To achieve these blistering speeds a radically new computer architecture
> must be used that has 1,048,576 processors. IBM will develop a new chip
> that contains 32 processors as well as lots of on chip memory but has
> only 57 machine level instructions compared with about 200 for most RISC
> machines.
Yes!!! They are going to processor in memory (PIM). You AI folks should be
wetting your pants with excitment. Think about it from the architecture
standpoint -- this means they can have a slightly different architecture
the processors only have a few instructions but effectively function
as neurons. By going to PIM, they are solving the interprocessor bandwidth
problem which is the *real* bottleneck to getting brain-equivalent computers.
> A great deal of effort will also go into making an operating
> system that is more fault tolerant than anything now in existence, if one of the
> processors or even an entire chip malfunctions the supercomputer will not die,
> it will just slow down very slightly.
As is true in the brain as well...
My calculations for the PIM approach give us brains on a desk sometime
between 2007 and 2015. Figuring 32 processors/chip, 1 million processors
works out to 32K chips, figuring 16 chips per board works out to 2048 boards.
It isn't going to fit on my desk but it will fit in a couple of racks.
For a first generation it isn't bad.
If they price this at $500/chip (processor costs), that works out to $16 Million
(big pharma & government only), but if they drop the price down to $25/chip
(current memory costs), that works out to $819K, still a little expensive
but universities and even small companies could afford one. If the price
falls by another order of magnitude ($80K; Korean or Taiwan fabs offer knock
off designs perhaps), then we are talking a range where the transhumanist
collective could buy one. If it comes down by two orders of magnitude ($8K)
a lot of companies will be working very very hard to replace wet brains
with dry brains.
It will be interesting to see what the power requirements are (after all
its mostly memory right?). Figuring 1W per chip (bet thats low), that
means I need 32K W, which at 120V works out to 273A. Damn, that exceeds
my current home electrical service, gotta call the electrical contractor
now. While I'm at it I better get the A/C guy too.
If anyone gets URLs with more technical details, please send them to me.
Thanks,
Robert
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:05:59 MST