Re: Parallel processing

[David Fotland <fotland@xxxxxxxxxxxxxxx>]

I am an expert in CPU design, as my job at HP is to architect HP's
Unix multiprocessor servers.  I was one of the inventors of the
PA and Merced isntruction sets, and I have advance knowledge of
HP's and Intel's CPU roadmaps.

On-chip gate delays are now around 0.1 or 0.2 ns, well under 1 ns.
Many companies are working on off chip signalling at over 1 Gbps
(less than 0.5 ns rise times).  CPU clock frequencies will continue
to rise at historical rates until at least 2002.  And when the Merced
instruction set comes into the mainstream PC, there should be another
near factor of 2 speed increase from that.

Highly parallel computers are still interesting, and in evaluating a go
position, one must read out many independent fights that can easily be
done by separate CPUs.  Michael Reiss used a parallel computer for his
go program about 11 years ago.

I'd like to use the new 1 million gate Xilinx gate array to do some
really fast custom hardware for go.  If we put 4 or 8 of these on a PCI
card, we can get parallel processing for go in a very small box.

David Fotland

At 07:34 PM 3/19/99 EST, Compgo123@xxxxxxxxxxxxxxxxx wrote:
>I'm not an expert in semiconductors, but do have some knowledge about
>electronics. I may be wrong, but it seems the CPU clock rate is appraoching
>the limit of transition time for silicon based IC's. I'm not aware of any
>circuit that has a rise time below 1 ns. Thus, it seems parallel
processing is
>inevitable including in computer Go.
>
>In this post I'd like to discuss the possibility of building a highly
parallel
>(with 10 CPUs or more) computer in one's basement. At present market price,
>one can get 10 266 Mhz CPUs for under $1000. One can get a sophisticated( the
>one for the parallel computer) printed circuit board made for several
hundred
>bucks. Digital and analog electronic design and simulation softwares, and
>printed circuit board design  softwares are widely available. Thus, the
>question is how do we design such a computer. Another problem is if there are
>software available for such a parallel computer. Software is a far more
>serious problem than the hardware because of the sheer man-years it needed to
>develop. I think this problem can be solved by design a operating system
level
>controled parallel computer.  One can build a front end that is compatible
>with available operating system (Windows98, Linux) and the compiler (i.g.
>VC++). Then, the front end communicates with rest of the system through
one or
>more high speed address ports. Or directly through the bus, just acts like a
>plug in board. Such a computer can easily increase the processing speed by a
>factor of 10 or more. The alpha-beta search in computer Go is very suitable
>for parallel processing. A commercial computer with eqivalent speed would
>costs several tens of thousand even more than one hundred thousand
dollars. On
>the other hand I think  above mentioned computer can be built for under
$4000.
>
>
>Dan Liu
>
>