[nadador]

I am curious if anyone here has experience with using the top end Power Macs vs. other similarly priced systems for doing scientific computations. I'm referring to a comparison such as the dual 2.7 G5 system vs. a dual Xeon Dell system (or dual Opteron HP or Sun workstation, or even a Sparc based Sun machine). We are currently looking to buy a machine to be used almost exclusively for running our in-house FORTRAN code and want to spend on the order of $5000 or less (no monitor). The computations are all about processor power, with relatively small RAM requirements, and data storage is not really an issue. Also, we have no need to do any visualization so graphics are a non-issue, and we are comfortable using OS X, Linux, or Unix.

Basically, what I've found is that the clock speeds of the AMD and Intel chips are faster and they have more L2 cache (especially the Xeon), but the FSB of the G5 is great. However, the 3.6 GHz clock speed of the Xeon and its 2 MB of L2 cache (large enough to make a difference with our small memory requirements) seem pretty good, so right now I'm leaning towards a dual Xeon machine from Dell. Does anyone notice something I'm missing or have thoughts as to why the G5's (or another system) would be better?

[OreoCookie]

You have to be more specific about the application. It depends very much on that plus what kind of compiler works for the code. How does it scale on several cpus? How sensitive is it to RAM throughput?

If you are interested simply in high linpac scores, the G5 is certainly a good choice (Mare Nostrum, Europe's fastest super computer will be built around it).

[Markarian421]

Scientific users I know that use a lot of FORTRAN code seem to prefer dual opteron machines (even ones with much slower clock speeds) over the Xeons, although I know a couple of them have dual G5's for their desktop machine. I don't know if they have done any head to head comparisons. If you want I can bug them about it.

[mousehouse]

I think IBM makes a Fortran compiler for the G5, XLF. The C-compiler xlc is reported to produce much faster code than the gcc compiler, at least about a year ago when i looked into this...

see:
http://www-306.ibm.com/software/awdt...x/xlf-mac.html

[OreoCookie]

GCC 4.0 seems to have caught up with xlc.

What version of Fortran do you need? f95, right?

[mousehouse]

offtopic:
is gcc 4.0 included in Tiger or Xcode 2.x ?

[OreoCookie]

Yes. Make sure it is selected when you customize your installation of XCode, I'm not sure if it is included by default.

[nadador]

The specific algorithm is a dynamic finite element code, with something like a Runge-Kutta algorithm for the time dependence. We haven't finalized it yet at all, so sorry if the details are vague. As far as scaling to multiple processors, it likely won't scale well, if at all. I'm mainly interested in multiple processors for running simultaneous simulations.

I have thought some about the compiler, and I think what we will need is f95, though due to a significant amount of legacy code I may use f77 (or atleast fixed format with f95). Either way, I know that Intel makes a compiler that can optimize for their processors, and the same goes for IBM and the G5 (though that compiler adds another $1000 to the price of the computer, whereas the Intel compiler I think is only $500 and g77 is free). Also, I know Sun has a specialized FORTRAN compiler for their computers. However, I don't know of any similar compiler for AMD chips, which is a drawback to me.

As far as how sensitive it is to RAM throughput, I can't test it yet as the code is not finalized. We estimate a couple hundred MB of RAM will be the simulation's maximum total usage, but every time step will require an update of all the data. Thus I'm guessing RAM throughput will be important. That's why the G5's high FSB is enticing, but since it runs the same speed RAM as the other systems (as far as I remember) perhaps that won't be as great as it sounds?

Markarian421, it might be helpful to me if you could ask why the Opteron's over the Xeon's, and do they use the dual G5's for computations (or is that just their preference for a desktop)? Just a short synopsis would be very helpful, no need to bug them too much.

[OreoCookie]

GCC now supports f95. In any case, you would have to add money for compilers if you want to get optimal performance. Intel compilers will also produce optimized code for AMD cpus, although you might have to disable cpu checking, but they will work (faster).

According to Germany's c't magazine, gcc 4.0 is slightly slower than xlc/xlf, but it is included on every MacOS X CD (you have to install the developer tools separately).

Also, your current budget (~ $5000) suggests that you probably won't be able to build a dual-core Opteron system as each cpu is already quite expensive (the cheapest ones goes for around 1000 € a piece plus RAM etc.). Opterons will be faster than Intel CPUs, in particular for memory-intensive apps

I would probably narrow it down to either a dual G5 or a dual Opteron box by – say – Sun. The G5 will (probably) be able to keep up with a dual Opteron box, unless everything fits into 2 MB or cache

[mousehouse]

If you get a PowerMac or SUN the Operating System is included. This usually is not the case with vendors selling "generic" Opteron or Xeon systems. Don't forget to include that in your costs.

Considering that the estimated performance of the G5's and Opterons will be in the same league I'd think that with a budget of $5K the PowerMac does get you a lot of mileage for the money. You get all the benefits from the Mac in your environment and ease of management. That will warrant the possibly small performance gap between the dual-G5 and a dual-Opteron.

[hypnognosis]

IIRC an Opteron with the intel compiler (assuming you force it to compile as a genuine intel) will give you the best performance. Sun also makes some very nice dual opteron systems with incredible I/O performance. The G5 with GCC 4.0 is also very nice if your doing BLAS. Unfortunately, if you want some nice low latency disk storage (I'm sorry a raptor doesn't cut it) performance you need to spend an additional $1000 on a reasonable scsi subsystem. G5's also don't come with ECC sdram, which I have always considered favorable for *nix workstations.

At the end of the day though, it shouldn't matter too much which system you use. Just ensure you choose a vendor that uses quality components and complement the system with the best available compiler.

[OreoCookie]

Originally Posted by hypnognosis

IIRC an Opteron with the intel compiler (assuming you force it to compile as a genuine intel) will give you the best performance. Sun also makes some very nice dual opteron systems with incredible I/O performance. The G5 with GCC 4.0 is also very nice if your doing BLAS. Unfortunately, if you want some nice low latency disk storage (I'm sorry a raptor doesn't cut it) performance you need to spend an additional $1000 on a reasonable scsi subsystem. G5's also don't come with ECC sdram, which I have always considered favorable for *nix workstations.

At the end of the day though, it shouldn't matter too much which system you use. Just ensure you choose a vendor that uses quality components and complement the system with the best available compiler.

The XServes accept ECC RAM, although it's not a must anymore (but still nice if you have it). It quickly adds up on your bill, though. Ditto for SCSI drives (which might not be necessary).

It all depends on the specific application. For some tasks, you can even use standard 100 MBit ethernet and start via network (because everything is in the RAM) while for others, this just won't work.

Depending on the job, a G5 cluster node might also be worth a second thought. Although I wouldn't recommend such a setup to be used as a desktop, more of a cluster approach.

[piet]

The unix base is a definite plus but the data flow in the G5 cpu is higher than the p.c. base cpu