[Beowulf] Intro question

[Lux, James P]

On 12/5/08 8:17 AM, "Lawrence Stewart" <larry.stewart at sicortex.com> wrote:

> I've been to a couple of DE Shaw talks and I always come away puzzled.
>
> It's tempting to conclude that they are just smarter than I am, but
> maybe they are just wrong.
>
> My understanding is they are building a special purpose molecular
> dynamics machine because it will be far faster than a general purpose
> machine programmed to do MD.
>
> In principle this might work, if you get the problem statement right,
> and you can design and build the machine before the general purpose
> machines catch up, and you don't make any mistakes, and after it is
> built you can keep designing new ones.  In practice it always seems to
> take longer than you expected and cost more, and maybe that 7 bit ALU
> really has to be changed to an 8 bit ALU to keep the precision up.

If the machine is built of reconfigurable FPGAs, then such a change is
pretty quick.

If you have a basic hardware infrastructure, and just respin an ASIC, and
that's a bit more time consuming, but not particularly expensive. e.g. Say
it costs, in round numbers, $1M to do an ASIC.  That's 2-3 work years labor
costs, so in the overall scheme of things, it's not very expensive, in a
relative way.  If your overall research effort is, say, $20M/yr (which is
big, but not huge), then budgeting for a complete machine rebuild every year
is only 5-10%.  If that gives you a factor of 3 speed increase, it's
probably worth it.

Think about it.. You check out your design in FPGAs to make sure it works,
then do FPGA>ASIC and crank out a quick 10,000 customized processors, have
them assembled into boards, fire it up and go. There are all sorts of
economies of scale possible (if you're building 1000 PC boards, on an
automated line, it's just not that expensive.  For comparison, we regularly
have prototype boards made with more than 20 layers and a dozen or so fairly
high density parts (a couple Xilinx Virtex II FPGAs, RAMs, CPUs, etc.) and
all the stuff around them.  In single quantities, it might cost around
$15K-$20K each to do these (parts cost included).  If we were doing 100 of
them, so we could spread the cost of the pick-and-place programming over all
of them, etc., it would probably be down in the $5-10K/each range.  Get into
the 1000 unit quantities where it pays to go to a higher volume house, and
you might be down in the few hundred bucks each to fab the board, and now
you're just talking parts cost.

Consider PC mobos.. The manufacturing cost (including parts) is well under
$100.

Now consider using that nifty compchem box to go examine thousands of
possible drugs.  Get a hit, and it can be a real money maker. Consider that
Claritin was responsible for about $2B of Schering-Plough's revenue in just
2001. Plavix was almost $4B in 2005.  That ED drug that starts with a V that
we all get mail about was in the $1B/yr area, although its dropping. (One
article comments that when it comes off patent in 2012 that they'll see a
bump in sales:"Recreational use of the product could also be expected to
generate substantial revenues.")

In this context, spending $100M isn't a huge sum, now, is it.



Jim