On Thu, 27 Sep 2007 22:43:50 -0500, Tom
wrote:
I want to build a good desk top number cruncher. Primary usage is a C
based program that performs the mathmatical calculations used to map
out a topology of various parameters. I do not need high end graphics.
Is a motherboard such as the ASUS P5K3 Delux a good choice?
First, you should determine which processor architecture
performs best at your particular App. Until you know this,
you can't know if a particular board is the best choice
since it will only support one of several CPU possibilities.
For some uses, including general PC use, a Core2Duo is the
best median-priced option, but when talking about a special
purpose build, Core2Duo may not be the right choice. In
some cases, even the dreaded Pentium4 is a better band for
the buck, but are you looking for best band for buck or best
performance, IOW, what is more important the price or
performance if you had to pick one, and what is the budget
ceiling?
http://computershopper.com/reviews/a...erboard-review
The above motherboard accepts up to 8 gigabytes (4 x 2) of DDR3 memory
at up to 1,600 MHz.
Ok, but how much memory do you *need* for your special
purpose? Beyond a certain point, adding more memory than is
needed just raises cost and/or requires slower memory
timings that will degrade the performance of the application
by a few percent.
Considering Vista Ultimate 64-bit is able to support up to 128GB of
memory ...
http://www.crucial.com/kb/answer.aspx?qid=4251
... what motherboard/memory card combinations should I consider that
could go beyond the above motherboard's 8 gig limit?
Number crunching seldom needs these large amounts of memory.
Would you be surprised if the typical number cruncher didn't
even need 128MB with WinXP? However, I don't know your
particular application, that will certainly matter.
I am currently looking at Intel Quad CPU's ...
Why? Is your app quad core optimized? Most number cruncher
apps aren't even dual core optimized. Generally
(historically) someone looking for good number crunchig
would want to determine if floating point performance is
important, what L2 cache size their working code fits
within, and within these limitations, pick the highest IPC
core and highest clockspeed core possible. When doing a
specific thing, the generalized web review benchmarks fly
out the window. It really matters a lot what the specifics
are of your particular task.
but understand that for
pure number crunching tasks there are other issues (i.e. L2 cache)
that enter into the performance calculation. Perhaps I should consider
an AMD configuration?
You should seek benchmarks of your specific app, or find a
roughly equivalent task keeping in mind issues of floating
point significance (or not), cache size, clockspeed, etc...
then seek benchmarks for that equivalent app on different
CPU designs.
A dream set-up would be a multibladed server with lots of cores
working independently ...
Really? Can your code accomplish this? If it can't, some
kind of exotic cooling to run fewer cores and insane speeds
would be more effective.
but for now a quad core desktop would really
be stretching my available funds. The ultimate would some research
super computer!!
Maybe, maybe not. Some tasks depend more on a fast memory
bus, and to make use of parallel processing you would
require code written to do it. IOW, on some supercomputers,
some apps will run slower than on a modern purpose-optimized
desktop.
Looking into DDR3 memory modules and latency issues
http://www.kingston.com/hyperx/products/ddr3_faq.asp
---------------------------------
Below is from above site:
Q. What latencies will standard DDR3 DIMMs support?
A. JEDEC DDR3 specifications define standard DDR3 CAS Latencies of 7,
9, and 10:
-1066MHz DDR3: CAS 7 (7-7-7)
-1333MHz DDR3: CAS 9 (9-9-9)
-1600MHz DDR3: CAS 10 (10-10-10)
----------------------------------
Seems the higher latency chews up a lot of the higher speed
performance?
I am starting to suspect your post is some kind of
kid's-dream, that you are not serious about really building
a system. You write about PC parts then DDR3 and
supercomputers. If you can divide the work like that, how
about just building separate budget-effective systems.
Three $1,000 systems will have higher performance than one
$3,000 system if you can do this.
Any advice for building a good number cruncher will be much
appreciated.
I will greatly value the input from those with 64-bit experience.
Why? Do you see a benefit from a 64 bit system? Your post
is far too light on details of your applications
optimizations and requirements.
Currently I am stuck in the single core, 32-bit realm and am looking
for the fast lane.
If you have no reason to believe these more exotic things
will help, a 32bit single core may actually be the highest
performing solution. It depends entirely on exactly what
you're trying to do, down to the last detail. I'm not
trying to discourage an elaborate cutting edge system but so
far, no details have suggested a gain from these things.