hi,

i'm currently running on an athlon xp 2400+ running at 2300mhz with 1G ram. I'm writting my own fortran code to run my NS solver. I'm wondering how much speed improvement i can get if i changed my mothboard and cpu to something like amd64 3000+? Is there any way to estimate? Or is there some benchmarking software?

Thank alot!

www.spec.org Within a CPU family it's likely to give you a reasonable idea. Less good when comparing AMD to P4 or Itanium. It wouldn't make sense to upgrade to an A64 3000, that is already obsolete and bottom of the performance range.

Hi

My experience shows that (FDM with Cartesian grid):

i attain about 2 speedup when i use AMD-64bit (128kb L1 and 512 kb L2-cache) instead Intel 3000 (8 kb L1 and 500kb L2-cache) when code is compiled with 32 bit format and when i use 64 bit format compiling (needs 64 bit OS) i attain about 4 fold speedup.

general remarks:

cache hierarchy is one of the essential feature of CPUs, L1 is more important, then L2, as AMD's CPUs have larger L1 we expect better performance.

also recent CPUs are usually use hyper-thread technology (e.g. intel 3000) for processing simultaneous jobs, so for acheiving maximum performance (100 % CPU usage) in such CPUs it is essential to use multi-thread programming.

also by increasing data locality and changing order of data access (such as mesh reordering) it is possible to decrease cache miss and so increasing performance considrably (up to 5 fold based on literature).

Good Luck.

Stating that AMD64 3000+ is obsolete is absolute nonsense. If we compare that market segment there is no CPU that beats it in price/performance, also with a few tweaks it can easily reach performance of much more expensive CPUs.

If you want to wait, Intel launches Conroe soon and amd will soon upgrade to ddr2 memory capability. This might or might not give a real boost to performance (most probably in the intel area), what is certain is that old motherboards will not be compatible with the new technology, hence waiting might be prudent.

My personal preference if I was on a tight budget would be to buy a single core opteron for socket 939 and some quality memory and then overclock it (takes some knowledge though). They are known to have an extreme overclocking potential.

As for performance gain, the only way to know for sure is to benchmark your code on your current setup vs the new setup since performance is not only cpu dependant but also memory and also (to some extent) other hardware.

>>> It wouldn't make sense to upgrade to an A64 3000, that is already obsolete <<<

An AMD64 3000 Obsolete?! What do you mean?

Perhaps "obsolete" is a slightly strong description, but a 3000+ is a long way from the fastest CPU's around today. If you are doing CPU-intensive work like CFD the priceerformance ratio of the CPU itself is not a particularly useful figure. Time costs money (lots of it), so within reason you want to get the fastest CPU that is readily available.

hmm.... i 've just tested a simple simulation from CASE lab research. It happens that my Athlon Xp 2400+@2200mhz with 1gb ddr400 ram is on par with the athlon xp 3000 and it's just 15% slower than a Opteron 242 1.6 with 2gb ddr333 ram.... is that possible?

i'm thinking of getting a Opteron 144 which still seems affordable. But if it's less than 50% improvement I don't think it's worth it.....

I have raced my athlon against my 64 (both 2Ghz) and the 64 beat it when doing the same Fluent model. Though, other things definitely come into the picture, so I really don't give any credit to the race. Overclocking can give you a little more kick, but if you don't have the proper cooling set-up (not a heatsink/fan) then you aren't really going to get much performance. MSI's motherboards have an applet that lets you overclock from Windows instead of the bios which is nice. And, it will tell you when your OS is becoming unstable. The applet will slowly step up the FSB until stability issues are encountered. You can also adjust V-core in the applet (not just the FSB). I used to work for a company called Kryotech that developed the first 1GHz desktop computer. It was done with some serious overclocking of an AMD Thunderbird chip that was cut for a higher multiplyer, and the cpu was cooled down to a chilly -40C via a regrigeration system.

Charles:

I really don't see your point here. So if money is not an issue you'll probably not look at mainstream CPUs and setups at all. Zonexo wanted to upgrade an old amd system with one CPU, not build a www.top500.org setup. That's my impression anyway. I might be wrong....

Andrew Hayes:

I run simulations on an overclocked system with stock cooling, with no minor gain. A 3200+ OCd at 2.5 GHz with memory going 1:1 in 2-2-2-5, and that's an old winchester. I'd say that's a Major gain!

Btw it's a DFI nf4 motherboard and twinmos bh5 memory.

At work I happily use our dual Xeon workstations, but my home setup is easilly a winner in price/performance. No question about it....

Ford Perfect:

well, when we overclocked the AMD chips we would take an 1200Mhz chip with a multiplyer of 9 and take it to 2.1Ghz with a multiplyer of 15. That is serious overclocking, which you can't do with a fan. The chip would overheat before it even made it to the bios check. We could get 18 multiplyers on some of the chips, but stability became an issue. But, even your increase in MHz on an AMD chip with just a fan is nice. You will have to let me know what kind to junction-temp you are seeing at the overclocking, and what kind of fan you are using. Also, are you upping the V-core? or is it strictly FSB?

I use a retail amd fan and arctic silver cooling paste.

I'm not sure what you mean by junction-temp, is it something used by MSI?

I'm currently at 2.3 GHz since I have stopped running simulations on my home computer (for the moment) and there is no need to stress the components. At 2.3 my CPU idles at 31 Celsius, the PWmic is at 36 C and the chipset is at 43 C. At the 2.5 GHz OC I don't remember the temperature, but it was not above 40 C idle. I have a CPU volt of 1.45 and Dram at 3.2 volts. If you are interrested I can make some adjustments and see what temp I get at 2.5 GHz, during Full CPU load. Multiplier is locked upwards so I use the standard 10x multiplier and 1:1 memory configuration (230 MHz FSB).

(for my intel system I use water cooling when I OC since the heat output by that system is much, much higher)

Are you doing your FSB adjustment in the bios? You can check you CPU temperature there. I am not sure what temperature your are referring to when you say 'CPU' temp; I am assuming it is chip temp since most modern applets display that. The bios will also tell you what your safe zone temp is. The problem with overclocking is if you you overclock too much and have to reset the bios, which is easy enough if you know what you are doing, but a hassle to get everything back to where it was before the crash. What we did when we were overclocking was step our FSB up and get into the bios when we started up and watch the CPU temp there - if it wasn't too hot we would up the FSB and restart. If I remember correctly, 60C is getting towards the danger level. The junction temp is the CPU chip temp itself (T-junction). The old AMD chips didn't have this and relied on a tiny thermocouple that was part of the MB to tell you the temp. The new chips have it built in.

Your finding of Opteron 242 performance vs. XP 2400 @ 2.2 GHz sounds about right. The 242 would only be much faster if you had two of them running a parallel job, which CASELab isn't. The 144 chip is only 1.8 Ghz. Remember that these CPU's all came out three years ago. Current top-end Opterons and A64's are around 2.8 GHz, so that is where your 50% has to come from. You * might * find a recently made 1.8 Ghz chip that you can push to 2.8 or so, but that is a different hobby, and there are no guarantees.

Money is always in issue. It's a question of what you spend it on. If overclocking is your game, by all means, have fun! But you can spend an awful amount of otherwise useful time dicking around with overclocking, and if you roast one CPU you've thrown away all the savings. It's worth experimenting a bit to see if the CPU you have is able to give you a little more (and 10% is often available with little risk), but once you start tinkering seriously, you are using time that should have been spent on doing some CFD work! Zonexo's problem is that he is trying to find a cost-effective way of getting a faster machine. From the description of what he has at the moment, he can't just pull out the CPU and replace it with an A64, so it becomes a MB & CPU upgrade. If he is doing that, well it probably makes sense to spend a bit more and get something closer to the leading edge of what's available than a 3000+. And he has already proved that the 3000 won't give him any extra.

As far as finding the fastest CPU's, unless the budget stretches to an IBM Power 5, mainstream is more or less the best you can get. I've used the Itanium II, and in the real world it doesn't live up to SPEC scores. So it really leaves you with mainstream CPU's as the best option for a CFD machine, but the cost of time can justify getting the top of the range. For the most part the only real difference between Opteron and A64 is that a 200-series Opteron is required for a dual setup. Otherwise, if it's a single CPU setup you want, I would suggest that the best option is one of the large cache A64 models.

Well in a sense, doing CFD at home on a low budget PC is a hobby too.

As for opterons, I was not talking about the 200 series. I don't think that they are available for socket 939 at all. What I meant was the 100 series and more specifically the low end, say 148 (somewhat similar to 3500+). If you push a 148 10% (which you stated as easy yourself) then you will probably have a faster system than a 150 system, unless you are really unlucky and unskilled. I don't know the price difference in you country, but there is clearly a gain, just by this extremely low overclock. I have seen numerous systems that push towards 2.8 and above. Then we have a serious gain, wouldn't you agree?

If you have an extreme ammount of money then going for the top model will of course always be the best option though.

I mean the three temperatures I stated. CPU, pwmic and chipset. Thermal specifications can be found at manefacturer's homepages.

I do the FSB and memory tweaks in the bios since DFI motherboards are specifically built for enthusiasts. The A64 tweaker can be used in windows to "soft" overclock an A64 system.

Many motherboards employ crash free bios today that reset your values if you fail during tweaks. And also gives option to save a specific setup.

Oh and the DFI board has prime95 build into the bios so that you can boot directly into a memory stress test. That is really nice if you want to test a setting for stability.

that is what I like about the MSI motherboards. I can overclock everything directly from windows. It will run a small stability test after each tweak and let me know if the system is still stable. Or, I can let it do it all automatically, and the computer will overclock everything to the safest level with the current set-up.

well, i'm into overclocking definitely. Right now, i can get the opteron 146 for around US$205 and 144 for US$179. of cos, there's still the motherboard which i must buy. It seems that it's not so difficult to reach 2.6-2.8 ghz. Assuming that is done, will there be 50 or higher % increase compared to the current AXP 2400+@2200?.

i've searched quite a few benchmark but there isn't any that seems relevant....

I think the best benchmark is to run your CFD code.

Otherwise a standard benchmark for CPU and memory is the super PI (easily downloaded, just google). Look around on OC forums and see what values on super PI your current setup gets and compare it to the setup you are thinking of obtaining.

Regards

FP

I recently bought an AMD Athlon 64 +3000 GHz processor and since then I've been questioning myself regarding the differences between Athlon 64 and Opteron CPUs, well, the following link supplied me some answers and I guess it would help you to make your decision.

http://www.devx.com/amd/Article/27340

Regards

Renato.