Re: Computer speed

[Konstantin <konst@xxxxxxxxxxxxxxxx>]

On Sat, 5 Dec 1998 Compgo123@xxxxxxxxxxxxxxxxx wrote:

> I recently cross over an article in a magzine that mentioned a little about
> super computers. It throw in some names like Cray T3E, IBM RS/6000SP, etc.
> Does anyone care to make some comment about today's (and in next 10 years)
> state of the art super computers? How do they compare with a Pentium 450 MHz?
> For computer Go, I guess the integer operation is more important. Thanks.
> 
> 
> Dan Liu
> 


In 1991, when I took a course in parallel programming at my university,
I had the opportunity to work on a Connection Machine CM200 with 8192
processors. Although it had a lot of processors, one has to remember the
fact that these were very simple 1-bit processors!!! 32-bit floating point
operations required 32 of these simple processors, thus leaving us with
8192/32=256 floating point processors.
Another important thing to remember is that not all programs are suited
for extreme parallelism. Thinking Machines (who built the CM200)  quoted 9
Gflops as a theoretical top performance on a fully configured computer
with 65536 processors. The one I was working on should therefore be making
at least 1 Gflops. That might be true for a finely tuned, specialised
piece of code that was written to use the computer at its full potential.
A real life application will almost surely have code in it that is not
parallelizable. For example, a program that I wrote, scored only 50
Mflops. That is 20 times less than what one might naively expect. I am
pretty sure that with a bit of hard work on fine tuning, I would have
gotten a lot more out of it, but reaching the theoretical peak performance
that the manufacturer quotes is a utopia.
To put the whole thing in perspective, one can compare with my newly
bought PII-450 128MB RAM, which scores around 50 Mflops of sustained
performance in a MATLAB program. The peak performance in my matlab program
is about 70 Mflops.
(True, true, not all programs can be written in linear algebra form, and I
admit that my program is highly parallelizable and would probably run
a lot faster on a Connection Machine than it does on my PC)
The point remains though that I can get the same order of magnitude of
performance on my 1998 $2500 PC that I got on the 1991 $1 million super
computer.
Needless to say that todays state of the art super computers would just
beat the shit out of my poor PC, but when I was buying my computer I
happened to forget my $1000000 at home...


If you do have a go code that is well parallelizable, and just happen to
forget your $1 million at home, don't despair... My guess is that cheap
FPGAs is a far better hardware choice than a super computer. Really messy
though in terms of both building the H/W and getting the S/W right. I
don't expect anything happening in go in that direction within the next 10
years. Ing prize is out of the question :-(




Merry Christmas...
/Konstantin