From: Richard Loosemore (rpwl@lightlink.com)
Date: Mon Aug 15 2005 - 13:40:26 MDT
Russell Wallace wrote:
> On 8/15/05, Richard Loosemore <rpwl@lightlink.com> wrote:
>
>>One time Director of Research at Star Bridge Systems, a dysfunctional
>>supercomputer company out of Salt Lake City.
>
>
> I heard of Star Bridge; there was a big flurry of announcements, then
> we never heard of them again. Are you at liberty to say what went
> wrong?
>
> - Russell
>
>
Oh, sure :-). But I will try to keep it short, since this only has the
potential to be relevant to SL4.
Background: Star Bridge Systems wanted to build supercomputers out of
field programmable gate arrays, because in principle this might allow
software to be implemented as circuits, with sets of gates dedicated to
individual algorithms, and with the circuits adapting fluidly as the
software ran. Theoretically, this could lead to speedups (for the right
kind of software, and especially for massively parallel AI software)
that involve silly numbers: factor of 10,000, for example.
The Technical Problems: (a) Mapping software to circuits is an
extremely hard problem. (b) Understanding the actual mechanics of
programming real FPGAs is more like art or voodoo than it is like
programming. (c) The FPGA chips are not currently designed to be
reprogrammed rapidly, on the fly. (d) If current FPGAs are used to full
capacity, the way that Star Bridge envisaged, they generate prodigious
amounts of heat.
The Spin Machine Problem: When SBS made their first announcement, they
talked as if they had solved or finessed all these problems. I was
invited to go work for them in the spring of 2001, but it took me six
months to finally realise that none of these problems had been solved,
that people had made some "exaggerations" about the technology that
bordered on outright lies, and that some of the problems were not even
*understood* by the founder or employees of the company.
[For amusement value only: The prototype, that they initially wanted to
lease for $26 million, could only be switched on for short periods
because one or other of its twin PSUs would burn out with a half-life
measured in hours. The original CEO had previously been a guest of the
government for a time. There were a couple of Gigs of ram on the
prototype boards, but all that time, until after I left, they couldn't
get the FPGAs to talk to it reliably. And when you ran your program on
the actual hardware, correct functioning depended on whether timing
problems would just happen not to be a problem. Sigh. ].
I remain hopeful that someone will buy the company, fire the founder and
get a gang of competent people in there to work on the actual problems.
As it is, the company survives on hype and the patronage of powerful
entities who throw money at projects they do not understand.
As I said above, this has the potential to be relevant to SL4 because a
reconfigurable supercomputer of this sort is urgently needed by the AI
community. The hardware problems (heat) may soon be solvable, with
radically new FPGA technologies that are over the horizon, and with the
right kind of internal design it might be possible to understand and
program the cells reliably. And meanwhile, the software engineering
community would have to do some lateral thinking on the issues involved
in real-world parallel programming.
You will all have to let me know if any of the issues mentioned in that
last paragraph are relevant to the SL4 list. Looks like no, from what I
see.
Richard Loosemore
This archive was generated by hypermail 2.1.5 : Tue Feb 21 2006 - 04:23:01 MST