Opteron or Athlon 64 FX for scientific workstation?

Typical job on a Pentium 4 takes a couple weeks in Matlab.
Which AMD 64 would speed me up the most? Plan to use a dual.

Thanks
Jim

On Sat, 08 Nov 2003 21:52:17 -0700, Jim Kroger
wrote:
Typical job on a Pentium 4 takes a couple weeks in Matlab.
Which AMD 64 would speed me up the most? Plan to use a dual.

Well, if you're looking for a dual-processor system, you need the
Opteron 2xx series chip. The Athlon64 FX (and regular Athlon64 for
that matter) are not dual-capable. Actually the Athlon64 FX is
exactly the same thing as an Opteron 1xx series processor, ie the
Opteron chips designed only for single-processor systems.

A nice dual-processor Opteron should help speed up your system nicely.
Unfortunately though, to the best of my knowledge Matlab is not yet
available for AMD64, so you'll be running in 32-bit mode for the time
being. The Opteron may very well be the fastest processor for the
task anyway, especially considering that the Itanium is not supported
at all and the Power4 doesn't seem to have any official support (not
to mention the fact that those are really expensive chips). Still, it
might be worthwhile comparing some numbers between the Opteron and the
Xeon for work as close as possible to what you do. With tasks running
a couple weeks, a few percentage difference means a lot and is worth
spending a little bit of time investigating things.

Heck, it might even be worth throwing a quick glance at how the
dual-processor Apple PowerMac G5 systems run. The performance of this
chip can be somewhat up and down as compared to PCs, but at least for
some situations it can be a much faster chip than any x86 chip out
there.

-------------
Tony Hill
hilla underscore 20 at yahoo dot ca

"Tony Hill" wrote in message
.com...
Well, if you're looking for a dual-processor system, you need the
Opteron 2xx series chip. The Athlon64 FX (and regular Athlon64 for
that matter) are not dual-capable. Actually the Athlon64 FX is
exactly the same thing as an Opteron 1xx series processor, ie the
Opteron chips designed only for single-processor systems.

Aren't all Opterons equipped with the three Hypertransport channels? How
exactly do they stop a lower-end Opteron from acting as a full
multiprocessor-capable Opteron?

Yousuf Khan

"Jim Kroger" wrote in message
...
Typical job on a Pentium 4 takes a couple weeks in Matlab.
Which AMD 64 would speed me up the most? Plan to use a dual.

Thanks
Jim

Would a better performing BLAS implementation improve your processing time?
Probably the best one is:

http://www.cs.utexas.edu/users/flame/goto/

"Yousuf Khan" wrote in
message
.cable.rogers.com...

Aren't all Opterons equipped with the three Hypertransport channels?
How
exactly do they stop a lower-end Opteron from acting as a full
multiprocessor-capable Opteron?

By not "pinning out" the additional channels?

By not testing, and therefore not guaranteeing, full multi-processor
performance?

These are the first two ways that immediately come to mind. There may
be others, short of using a lazer to blow some links - that would work,
too.

Felger Carbon wrote:
"Yousuf Khan" wrote in
message
.cable.rogers.com...

Aren't all Opterons equipped with the three Hypertransport channels?
How
exactly do they stop a lower-end Opteron from acting as a full
multiprocessor-capable Opteron?

By not "pinning out" the additional channels?

By not testing, and therefore not guaranteeing, full multi-processor
performance?

These are the first two ways that immediately come to mind. There may
be others, short of using a lazer to blow some links - that would work,
too.

As an aside, these are the prices for various 1.8 GHz Opteron's as found
on Newegg's web site.

Opteron 144 OEM $ 283
Opteron 244 OEM $ 441
Opteron 844 OEM $1260

Just as a comparison, Intel has
Itanium 1.4 GHz 1.5M L3 $1172
Itanium 1.4 GHz 4.0M L3 $2247

So Intel charges a lot of money for each extra MB of L3, and AMD
charges a lot of money for each extra (functional and validated)
ccHT link.

--
davewang202(at)yahoo(dot)com

On Mon, 10 Nov 2003 06:53:40 GMT, "Suen Lee"
wrote:

"Jim Kroger" wrote in message
.. .
Typical job on a Pentium 4 takes a couple weeks in Matlab.
Which AMD 64 would speed me up the most? Plan to use a dual.

Thanks
Jim

Would a better performing BLAS implementation improve your processing time?
Probably the best one is:

http://www.cs.utexas.edu/users/flame/goto/


That's a good suggestion, but if you are doing jobs that take weeks,
you probably want to consider some kind of clustering solution, which
need not be expensive.

How to do it depends alot on the details of the problem you are
solving. Google

matlab cluster

and

matlab "parallel implementation"

The matlab licenses will probably cost more than the hardware. :-(.

RM

On Mon, 10 Nov 2003 06:15:40 GMT, "Yousuf Khan"
wrote:
"Tony Hill" wrote in message
t.com...
Well, if you're looking for a dual-processor system, you need the
Opteron 2xx series chip. The Athlon64 FX (and regular Athlon64 for
that matter) are not dual-capable. Actually the Athlon64 FX is
exactly the same thing as an Opteron 1xx series processor, ie the
Opteron chips designed only for single-processor systems.

Aren't all Opterons equipped with the three Hypertransport channels? How
exactly do they stop a lower-end Opteron from acting as a full
multiprocessor-capable Opteron?

Three hypertransport channels, yes. Three *cache coherent"
hypertransport channels, apparently no, or at least that is my
understanding. What I've seen suggests that hypertransport has to
operate in a special cache coherent mode to work as a multiprocessor
link. Otherwise it's a fairly dumb (but effective) point-to-point
link.

Of course, that could just be a bit of marketeering getting thrown
into the mix.

-------------
Tony Hill
hilla underscore 20 at yahoo dot ca

Robert Myers wrote in message . ..



How to do it depends alot on the details of the problem you are
solving. Google

matlab cluster

and

matlab "parallel implementation"

The matlab licenses will probably cost more than the hardware. :-(.


There are a few research efforts for parallelizing Matlab, but Matlab
has abandoned their work in this vein. Current versions are
single-processor only. Of course you can fork as many as you want, and
a dual will assign each processor to an instance of Matlab, but that
is not always helpful for a given analysis. My approach on duals has
been to just use each processor to run an analysis independantly.
Still takes a couple weeks but at least two get done instead of one.

Jim

On 10 Nov 2003 17:49:51 -0800, (Jim Kroger)
wrote:

snip

There are a few research efforts for parallelizing Matlab, but Matlab
has abandoned their work in this vein. Current versions are
single-processor only. Of course you can fork as many as you want, and
a dual will assign each processor to an instance of Matlab, but that
is not always helpful for a given analysis. My approach on duals has
been to just use each processor to run an analysis independantly.
Still takes a couple weeks but at least two get done instead of one.


I'm intrigued. This is not the place for a detailed discussion, of
course, but I'd be interested in learning how much of the
single-threadedness of your problem is matlab and how much is
intrinsic to the problem, if there are any details you'd care to
share.

RM