Best 'budget' processor for scientific applications?

[mmganga]

Hi,

I am looking to see whether I can get any noticeable gain from my current setup by upgrading it to something else...

I have a NF-2 mobo with a Mobile-2500 CPU which overclocks to 2400 MHz (quite stable, I haven't had a single crash!!!)...this is on air cooling.

I run mostly scientific applications with large matrices and inversion operations on them...the applications are 32-bit (for now at least) so 64-bit won't help much. There's no substitute for CPU speed and efficiency here!!!

From what I benchmark with ScienceMark I see my levels are around the XP3200+ levels, and definitely higher than P4-CPUs...

So I'd be looking to see what (and IF) I should take any steps for upgrading this to a better setup. Note that a 5% improvement won't help me much, I'm looking for something like 25% at the least.

Can you guys give any recommendations what it would take to get more speed out of this CPU or get another CPU which will go much higher in MHz speed (ie. 25%)??? AND do it all on a reasonable budget (ie. at most watercooling, nothing crazy like vapochill setups and such)...

I know I could benefit from the FX-53/55 series of CPUs but the price is not worth the time savings...note that in ScienceMark even the FX-53 is not that much faster than what I have right now (maybe 7-8%???). Now a FX-53 overclocked to 3+ GHz is a different story but that will run $$$, am I right?

THANKS!!

 

[nemiAMD]

Hey, If it's not too sectret why don't you complie a little .exe of you most common routines and ask a few people here to give you benchmarks on their systems?

If you think that memory bandwidth might have somehting to do with your performaqnce on these maths routines then there seems to be lots of people here with (dual channel) RAM going greater than PC3200 speeds....

 

[mmganga]

hey that's a great suggestion, I will look into it!!!

There's no secret what I'm doing, although it involves using both my own numerical codes and commercial software (licensed to my school)...it's basically

But I'll put a bunch of codes together and upload it for people to benchmark it...it would be of great help!!!

THANKS!!

 

[miketheidiot]

Originally posted by: mmganga
Hi,

I am looking to see whether I can get any noticeable gain from my current setup by upgrading it to something else...

I have a NF-2 mobo with a Mobile-2500 CPU which overclocks to 2400 MHz (quite stable, I haven't had a single crash!!!)...this is on air cooling.

I run mostly scientific applications with large matrices and inversion operations on them...the applications are 32-bit (for now at least) so 64-bit won't help much. There's no substitute for CPU speed and efficiency here!!!

From what I benchmark with ScienceMark I see my levels are around the XP3200+ levels, and definitely higher than P4-CPUs...

So I'd be looking to see what (and IF) I should take any steps for upgrading this to a better setup. Note that a 5% improvement won't help me much, I'm looking for something like 25% at the least.

Can you guys give any recommendations what it would take to get more speed out of this CPU or get another CPU which will go much higher in MHz speed (ie. 25%)??? AND do it all on a reasonable budget (ie. at most watercooling, nothing crazy like vapochill setups and such)...

I know I could benefit from the FX-53/55 series of CPUs but the price is not worth the time savings...note that in ScienceMark even the FX-53 is not that much faster than what I have right now (maybe 7-8%???). Now a FX-53 overclocked to 3+ GHz is a different story but that will run $$$, am I right?

THANKS!!

A 2.4 a64 should whoop a 2.4 Axp.

 

[Ackbar]

I was looking into the same thing. My work also requires a lot of computational power, but really it's forced me to program better. My suggestion if you really need brute computational power would be to make one of these guys...
http://www.xtremesystems.org/f...showthread.php?t=50314

I was considering one for a while, but decided against it at the end since my funds are limited (still using a 6 year old computer that wasn't fast when I got it!) Just program better, you'd be surprised how much you can gain in speed if you just try and make appropriate approximations and such. Good luck

 

[mmganga]

Well it's not as simple as that, if it were my own programming maybe it could be improved (although I am no great programmer by any means ;-)...

Problem is that I'm also using some commercial codes which build large (somewhat sparse) matrices of ~200,000 x 200,000 complex elements and then try to solve them (Finite Element Method -- FEM)...

In that case the matrix size is maybe 1 Gb in memory and ~2 Gb swap, but A LOT of time is spent in CPU time, although swapping is also a concern.

For swapping there's no solution other than memory increase and I'm working on that, but still the CPU is also a bottleneck...

AS far as the Xenon suggestion, those CPUs seem to run to ~2.8 GHz, but they are inefficient as far as per-cycle instruction goes...I mean AMD 2.5 GHz beats a P4 3.2 GHz and so on...so I think in the end the question is whether to upgrade to a better AMD processor.

Any other ideas guys?

THANKS!!!

 

[Zebo]

A68 2800s

 

[Ackbar]

Well, if you're not running your own code, then you probably won't be able to take advantage of the second CPU (non-multithreaded applications). If you don't mind me asking, what are you trying to compute?

Finite element analysis does take up a lot of memory... you could always diskswap but that's highly inefficient. Sounds like you just need more RAM. As far as I know AMD 64's don't really need that much memory bandwidth, but other processors seem to like it a lot especially non-dual channel. What types of improvement are you looking for half as long? 1/3 as long?

 

[Accord99]

Originally posted by: mmganga
From what I benchmark with ScienceMark I see my levels are around the XP3200+ levels, and definitely higher than P4-CPUs...

ScienceMark is a poor benchmark of Matrix calculations. It would depend on what you're doing, ie using the well-optimized Goto BLAS library for dgemm, Prescott-based processors are more efficient clock per clock than an Opteron.

 

[mmganga]

Zebo, you mean A64 2800's??? I don't know what A68 is!!

Ackbar: I use HFSS (www.ansoft.com) for Electromagnetic solving of 3D structures...pure FEM code...the structures can be anything from small antenna designs to PCBs (well more like transmission lines) and 3D waveguiding structures...

Obviously the larger the physical size of the structure the more time it takes to solve it...

 

[Ackbar]

Interesting... I'm guessing it does a lot of solving of Maxwell's equations then using matrices for the fields and what not. With the main computational aspects being integration... well, I guess the application doesn't matter as much as what your budget is. Are you looking for an entirely new system or just CPU upgrade?

 

[mmganga]

well mostly CPU upgrade, or basically to see what's the best way to improve my sim times...

 

[Vee]

My own approach to matrix & branch heavy work was s754 A64 3400+ non-OC on a fairly simple MB (EPoX 8KDA3J). Going from (often throttling) P4s that have been a dissappointment, ever since the PIII, the experience was very satisfying. However, I'm convinced a large part is due to code not being well suited to the P4. So since the A-XP is already a brute at '87 FP math and branches, you might not benefit so much, or at all, going from a XP at 2.4GHz. A64's seem generally very good at these kinds of apps, but maybe you shouldn't expect too much? Best would be if you could benchmark your app, prior to any decision.
(Otherwise, since you're an overclocker, forum people are very enthusiastic about s939 3200+ and 3000+.)

 

[Vee]

Originally posted by: Accord99

Originally posted by: mmganga
From what I benchmark with ScienceMark I see my levels are around the XP3200+ levels, and definitely higher than P4-CPUs...

ScienceMark is a poor benchmark of Matrix calculations. It would depend on what you're doing, ie using the well-optimized Goto BLAS library for dgemm, Prescott-based processors are more efficient clock per clock than an Opteron.

If you have SSE2 optimized libraries, P4s are very good at simple stuff like matrix muls and adds. But I suspect that the P4 is a disaster at inversion. Likewise if you don't have optimized libraries.