home .. forth .. machineforth mail list archive ..

[MachineForth] Chuck's 25x and IBM's Blue Gene


----- Original Message -----
From: "Dirk Heise" <dheise@xxxxxxxxxxx>
Subject: [ColorForth] machineforth and colorforth info at Chuck Moore's new
website


> Hey great! Words from the master himself!
> And that 25-CPU chip sounds like what Intel
> wants to arrive at some of the next decades...
> If one of you in the USA starts producing it,
> i'll do the distribution in Germany!
>
> Dirk Heise

In the current issue of Wired (July 2001), there is an article about IBM's
Blue Gene machine, which will be based on chips which sound similar to
Chuck's 25x.  Blue Gene is planned to have ~32 CPUs per chip, which are then
combined into 36 chip-boards, then into 4-board towers, and finally into a
256-tower cluster.  Altogether that's around a million processors!

The IBM chips will have about 4 kilobits of embedded DRAM for each CPU.  And
the chips will be quite RISC-y, perhaps even MISC-y.  In Wired's words,
"Everything is stripped of unnecessary complexity: Denneau [of IBM] is
taking what's called reduced instruction set computer (RISC) ideas a lot
further than anyone else at IBM, constricting the number of different things
his chips can do to a bare minimum... 'we found an amazing thing... all we
ever saw used were the same 50 or 60 or 70 instructions' [in program traces
of other IBM chips]".  IBM expects each CPU to produce about a gigaflop...
about a petaflop for the entire machine.

OK, so Chuck can live with 1/2 to 1/3 that number of instructions, but it is
interesting to me that the biggest computer company in the world is going to
use very similar approaches to address one of the most complex computing
challenges ever: modeling of protein molecule folding.  While Chuck is ~15
years ahead, IBM has the resources to cut its own path.  It would be a shame
though if IBM reinvents Chuck's technology without considering and perhaps
adopting it for this computing "Grand Challenge."

Some possible differences between 25x and IBM's chips: the IBM will probably
be heavily floating-point oriented (with perhaps an order of magnitude more
transistors), and will be designed for multiple thread-style programming.
It appears that 25x will be integer-only and each of the 25 processors will
tend to be assigned a fixed-purpose per application, although this is not a
requirement.

Anyway, the Wired article had some other interesting info on why protein
folding is such a difficult challenge: it's a many-body problem for you
physics types out there... therefore you know there is no "closed solution"
for the motion of the particles (even in classical physics).  Add to it
quantum mechanics effects and it becomes intractable with simplistic
modelling approaches and today's machines.  More sophisticated modeling and
simulation software, along with huge increases in processing power are
highly desired in this cross-pollination of biology, quantum chemistry,
computer engineering, and computer science.

I imagine that someone besides IBM wants to build a Blue Gene-class machine
for this or another application.  With '25x' parts I would guess it can be
done for 10 to 25% of the cost.. perhaps much less.  Perhaps some bio-tech
company has $0.5 million to fund Chuck and another $10 or 20 million to
design and build a cluster!  Of course, they would also need some Forth
programmers to get their algorithms working! :-)
--
Mike Losh        http://members.home.net/~forthist/

------------------------

To Unsubscribe from this list, send mail to Mdaemon@xxxxxxxxxxxxxxxxxx with:
unsubscribe MachineForth
as the first and only line within the message body
Problems   -   List-Admin@xxxxxxxxxxxxxxxxxx
Main Machine Forth site   -   http://www.