Opteron Fortran (was Re: flood of bounces frompostmaster at systemsfirm.net)

Douglas O'Flaherty douglas at shore.net
Wed Oct 22 01:33:02 EDT 2003

Here's the short summary of Opteron compilers. When someone offers an 
AMD64 compiler, it typically may be used to create 32-bit or 64-bit 
executables as long as you are specific about which libraries you use. 
Any IA-32 compiler can create code and run on Opterons. Of course, 
32-bit executables don't get the extra memory either, even when running 
on a 64-bit OS, but sometimes a 32-bit executable might be what you want.

With SC2003 coming up, I expect we'll see a flurry of activity relating 
to compilers and tools. This information will likely be stale soon. 
Also, most of these have a free trial period, so you can kick the tires.

Intel compilers work great in 32-bit and can be run on a 32 or 64-bit OS 
natively. Performance and compatability is not an issue. For obvious 
reasons many of the benchmarks have been run using IFC.

PGI's first AMD64 production release was around July 5.  There is a 
limitation on objects greater than 2GB in Linux as a result of the GNU 
assembly linker, but the application can address as much memory as you 
can give it. Only a small fraction of the world has objects that large. 
I've only run into it with synthetic benchmarks. The gal coding is done 
and PGI is working on the next release. As for performance, since this 
was the first AMD64 fortran compiler to market, it was used in AMD 
presentations. You can see performance comparisons in Rich Brunner's 
presentation from ClusterWorld. It's on-line at 
(about slide 39 IIRC) There was a minor patch release near the begining 
of August. I suspect there is always someone finding flaws, but 
generally it's doing well.
NB: Saw Glenn's post re: PGI on SuSE v. RedHat. We've got it running on 
both. There were definately some fiddley bits to make it happy on 
RedHat, but I think they are documented on PGI's site.

Absoft had a long beta of their AMD64 compiler and went GA in September. 
I have no personal experience on it, nor do I know of any public benchmarks.

NAG worked closely with AMD on the AMD Core Math Libraries. They should 
know the processor well.

No experience with the Gnu Fortran or Lahey. I believe GFC to be AMD64 
functional. Lahey would only generate 32-bit code.

Your other question was about SSE2. Yes Opteron has complete SSE2 
support. I *know* PGI & IFC support it, I expect the others do as well.


Disclaimer: Among my several hats I am also in AMD Marketing. This is an 
unofficial response. No AMD bits were utlized in the creation of this 
email, etc..  If you want to talk about Opterons 'officially' you need 
to email me at doug.oflaherty(at)amd.com

On Mon, 20 Oct 2003 at 7:08pm, Robert G. Brown wrote

>> On Mon, 20 Oct 2003, Trent Piepho wrote:
>>> > I'm getting a flood of bounced messages from postmaster at systemsfirm.net, every
>>> > messages I've sent to this list has started bouncing back to me from
>>> > dan at systemsfirm.com.  I'm getting about ten copies of each one every other
>>> > day.  Is anyone else having this problem?
>> BTW, you (and of course the rest of the list) are just the man to ask;
>> what is the status of Opterons and fortran compilers.  I myself don't
>> use fortran any more, but a number of folks at Duke do, and they are
>> starting to ask what the choices are for Opterons.  A websearch reveals
>> that PGI, Absoft, NAG, Lahey, and perhaps others claim to have an
>> Opteron fortran, but rumor also suggests that a number of these are
>> really "beta" quality with bugs that may or may not prove fatal to any
>> given project.  Then there is Gnu.
>> Any comments on any of these from you (or anybody, really)?  Is there a
>> functional 64-bit Gnu fortran for the Opteron?  Does Intel Fortran work?
>> Do the compilers permit access to large (> 3GB) memory, do they optimize
>> the use of that memory, do they support the various SSE instructions?


Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

More information about the Beowulf mailing list