Hello CFD community ? My question is what you recommend to buy for 2.500$. Would you buy one UNIX workstation or several PCs? What is the optimum for this money ? Thanks

For $2,500 UNIX workstations are out of the question - they are much more expensive. If you prefer UNIX I would recommend you to buy a 1.7GHz P4 (the P4 runs Fluent very well!) with 512MB memory and install RedHat Linux on it. You should be able to get a decent 19" screen and a good graphics card (GeForce3) in the $2500 budget. Dell sells a nice budget-machine (Dimension 8100).

If you want to run large cases a nice alternative could be an Athlon 1.33 GHz machine with more memory (1 gig or so) - CPU and memory is cheaper for this config so you can buy more memory. The performance running Fluent on the P4-config will be about about 40% better than the AMD-config though, unless of course you need more memory than you can afford with the P4-config.

Thanks for your answer ? But what about two processor machine if I did not buy monitor ? I have also heard that P4 uses old technology and is not as good as it is promoted.Does AMD support two-processor platforms ?

AMD just released a SMP board and technology, but it is literally brand new and relatively expensive - approximately 500 bucks for one of the few boards available. I haven't heard any problems using the P4, so I would go that route.

~patrick

Until recently neither the P4 nor the Athlon was available in dual-CPU configurations. However, this is changing - dual CPU P4 machines were recently launched. Dell, for example, sells a nice new machine called Precision 530 which you can get Linux pre-loaded on. However, you will not get one of these for $2500, even if you skip the montior. A bare-bone dual P4 1.7 GHz with 512 MB RDRAM (minimum) will cost about $3500 I think. AMD is also just about to launch a dual-CPU Athlon version (code-named Palomino). I still haven't seen any machines based on it though - should be out in a few months I think.

Jonas,

Sorry to stick my neck out on a non-CFD topic, but do you have something to backup your claim of 40% better performance on the P4 1.7GHz wrt an Athlon 1.33 GHz. Is that true?

The general word around is that for the same clock speed the Athlon is 30% faster than a PIII on x86 code. And it is generally accepted that the P4 runs x86 code which is non-P4 optimised slower than the fastest PIII or fastest Athlon. See www.tomshardware.com for example, especially the mpeg-4 encoding articles. There is no reason to believe that CFD software would not behave similarily. Even if memory bandwidth was an issue, it would not make a 40% difference, maybe 10% at most.

So, unless Fluent compiles specifically for the P4, which I would think is unlikely because the compilers were just released recently, then something is amiss.

Dan.

Yes, I have hard benchmark-numbers comparing Fluent 5.5.14 on a 1.5 GHz P4 with RDRAM, an 1.33 GHz AMD Athlon with SDRAM and an 1.33 GHz AMD Athlon with DDR SDRAM. Several different cases were run. Here is a typical example:

<TABLE CELLPADDING=10> <TR><TD>Processor</TD><TD>Memory</TD><TD>Speed</TD></TR <TR><TD>1.5 GHz P4</TD><TD>PC800 RDRAM</TD><TD>1.33</TD></TR> <TR><TD>1.33 GHz AMD Athlon</TD><TD>PC2100 DDR-SDRAM</TD><TD>1.05</TD></TR> <TD>1.33 GHz AMD Athlon</TD><TD>PC133 SDRAM</TD><TD>1.00</TD></TR> </TABLE>

As you can see the 1.5 GHz P4 is here 30% faster than then Athlon. I agree with you that this is surpricing and before I saw these results I were also sceptical of the P4. Note also that the Fluent version used is the standard shipping 5.5.14, which is not in any way optimized for the P4. The benchmarks were run on RedHat Linux.

As a side note - the 1.7GHz P4 is probably now the fastest single-CPU Fluent machine available now, beating the most expensive workstations from HP, Compaq and SGI!

Check the SPEC comparisons. These are good benchmarks for Fluent and other computation intensive programs.

Floating point scores: P4 1.7 GHz is 598, Athlon 1.3 GHz is 414. Result: 44% faster for the P4.

Integer scores: P4 is 593, Athlon is 496. Result: 19% faster for the P4.

The Athlons are faster for many other programs (like those shown at tomshardware.com and other GAMING ONLY sites), but for pure computations P4 should be from 19% to 44% faster.

The spec numbers are not really relevant here since the P4 numbers were produced by Intel using P4-optimizing in-house compilers that make use of the new SSE2 extensions. No commrecial CFD codes use this as far as I know.

For the moment, I agree that there are probably no optimized SSE2 CFD codes. However, I can see no reason why these codes won't be available soon. But, for now we will not see the benefit of these new commands. From all the reviews that I have seen the SSE2 optimizations speed up the processor by roughly 10%-15%. Lets ASSUME this is also correct for Fluent calculations. Dividing my previous post numbers by 1.15:

Floating point: P4 is 598/1.15=520, Athlon is 414, Result: 25% faster for the P4. (This is very similiar to the 26% speed difference you reported earlier.)

Integer result: P4 is 593/1.15=516, Athlon is 496, Result is 4% faster for the P4.

Conclusion, until Fluent is optimized using SSE2, I'd expect 4% to 25% faster for the top P4 compared to current top Athlon for CFD codes. True, the memory cost of the RDRAM is excessive, so there is a better Performance/Price ratio for current Athlon computers. However, I was trying to point out to Dan Williams that the P4 can give nearly the 30%-40% increase that you originally claimed. No benchmark is perfect (applicable to all programs), but the SPEC numbers are a great start, until Fluent updates their own benchmarks (they don't have any Intel processor faster than 733 MHz or a single Athlon listed).

First, www.tomshardware.com is one of the best sources of PC hardware reviews that I know of - certainly not to be dismissed with a simple "gamers only site" comment!

SSE2 optimizations can produce dramatic speed improvements for floating-point intensive applications like CFD, much more than the 10-15% you mention. For example, when Intel-engineers worked on improving the Flask MPEG encoding benchmark with SSE2 instructions they managed to increased the performance by a factor of 3.5! (see here).

There is one good reason why we won't see SSE2 optimized commercial CFD codes for some time - there is, as far as I know, no released compiler on the market yet which can produce this kind of code. Intel has released a beta-version of their compiler, but this was just recently and the beta-period extends over the summer (ends in September I think). CFD vendors are now starting to test it, but it will certianly take some time before they release a version compiled with this compiler - I guess that they will wait until the compiler is out of beta. When I asked Fluent about this a year ago they had no plans to produce a P4 optimized version, but I think that they are starting to change their mind now that they see how well the P4 runs the non-optimized Fluent version.

Your example of FlasK MPEG encoding could be quite misleading to some readers. The recompiled program did improve by a factor of 3.66 when using SSE2 optimizations. However, the P3 and Athlon which do not have SSE2 capability also improved by nearly a factor of 2 each. Note: tomshardware.com has an update to that link which claims that the Athlon would have improved by even more than a factor of 2 if the code was optimized for the Athlon. Thus I agree that SSE2 optimizations can improve performance. However the claim of a factor of 3.5 is quite misleading, since there was much more optimizing done than just adding SSE2 capability.

If Fluent would run 3.5 times faster just from a simple recompilation, then they certainly should have enormous pressure from its users to recompile. I wish it was true - simulations that usually take 1 week would then require only 2 days without a single hardware upgrade...

I appologize, I should have said "Gamers focused". Tomshardware has some very good advice in its reviews. However, I'd avoid taking the comments in the forums section to heart since the vast majority of them are gamers. For example, post a question asking if upgrading to a 3Dlabs® Wildcat II 5110® would help your performance. (I've tried; 95+% of the responses said that the card is worthless and its performance can be topped by even cheap, old videocards. That may be true for some games but the Wildcat is easily the fastest available for CAD...)

Sure, the 3.5 time improvement is quite extreme, and probably not represantative of what a CFD code would get. However, it clearly illustrates the performance difference you can get by optimizing code. Since the spec-numbers for the P4 were produced by Intel I assume that they used every possible tweak to make them as good as possible, like they did with the Flask MPEG test.

This is all quite interesting, but it still doesn't change the fact that the for the same clock speed the Athlon architecture is 30% faster than a PIII and only slightly slower (5-10% at most) than a P4. So, all AMD has to do is improve the architecutre and ramp the clock speed and they will have the lead again. The fact that intel is 30-40% faster than an Athlon by simply using 30% more clock speed does not impress me. (Note that I'm not a big Intel fan though).

Other than intels compilers, I don't know of a compiler that makes SSE2 optimised P4 code either, so until the compilers are officially released there's not a lot one can do. In the mean time Portland Group and the DVF compiler both produce optimised code for the SSE (on PIII) instructions, and implement cache prefetching (both PIII and Athlon). These optimisations add roughly 30% more performance to compiling without them.

Dan.

(1). I think the current PC speed is fast enough to do most simple 3-D cfd problems. (2). I bought my first Radio Shack computer for around US$2500.(it had one MHZ CPU in it) (3). Now the Clock speed is 1000 times faster. Between 8-bit CPU and 32-bit CPU, there is a speed up ratio of 4 to access the same length word. So, the total speed ration would be around 4000 times.(4). With 2500 dollars, I would buy two computers, a laptop (around 1500 dollars) and a desktop model (around 1000 dollars). (5). The speed of CPU is going to hit the limit soon. This is because the current speed is far too fast for the office word processing application, and the investment in research to increase the speed is also very high. (one of the Intel founder has just retired recently). (6). My feeling is: two GHZ is adequate for average CFD applications. For me, one GHZ should bring my 3-D code into less than one hour range. (my simple 3-D code using Pentium/100 took about one hour) (7). So, if you are still using the very slow and big commercial codes, you should start looking at the possibility of writing your own codes. (8). I am not writing cfd codes right now, because of the limited market. (9). With the current poor PC market, the speed of CPU will probably stay in the one to two GHZ range for some time in the future. From my point of view, it is adequate for most cfd applications, if you are writing your own codes for average applications. (this rules out reacting flows and multi-phase flows)

Dan:

Continuing your logic...

The mac G4 733 MHz is roughly the same speed as an Athlon 1.3 GHz, thus the G4 is 77% faster clock-for-clock (if only macs could ramp up the speed...)

The Intel Itanium 800 MHz is roughly double the speed of Intels P4 1.7 GHz, thus it is 400% faster clock-for-clock (if only Intel could ramp up its Itanium speed more than Intel ramps up its P4 speed...)

I could go on for a while. There is a major reason why Athlons aren't past 1.4 GHz - they produce enormous amounts of heat, so much that AMD can barely cool them down enough. Increasing the speed will produce more heat and thus burn up the processors. It is not simple for AMD to just ramp up the clock speed. If only there was a way of modeling the cooling of the processor to get around this limitation (hint CFD).

Of course AMD will shrink their die size (to produce less heat) as soon as they have the technology, but so will Intel - bringing us back to square one.

My point is that many companies can get similar results with vastly different clock speeds. So, comparing processors clock-for-clock is quite meaningless. The only resonable comparison is to get the top processor you can afford of each brand and compare those speeds (keeping in mind that the prices may vary by 1000%, even within the same company).

I assume Jonas's Fluent benchmarks are accurate; the P4 is 30% faster than the Athlon. However the P4 costs roughly 50% more than the Athlon. For the $2500 budget, a P4 may be out of Kim's price range. Dual Athlons are available now and may be just inside Kim's range (I think the motherboard price is exceedingly high). Dual P4s are out of the question, Kim needs at least $3500 if not $5000. Dual P3s are also just out of Kim's reach at $3000 for a good model. I've never seen a good SUN, SGI, etc... for $2500.

(1). More than a year ago, there was a PC with a dedicated liquid cooling system to cool the CPU for over-clocked applications. I don't know whether it was a prototype or a commercial system. So, CPU cooling is definitely an issue. (2). Such dedicated cooling system has long been used in the super-computers such as CRAY computer. The successful development of CRAY computer depended on the cooling system. (3). Such dedicated cooling system for PC is not going to be very practical, because of the power shortage in California. You have to realize that, with all the computing power available to NASA, the first hypersonic X-43A(?) testing was a failure in first few seconds as reported in news a couple of days ago. They must have been developing the so-called CFD codes for the last 15 years for such applications. (4). What you need is affordable computers and an outstanding person who understand the problem to be solved. CFD codes and faster computers are necessary, but the key to the success is "human brain".(which is always very hard to find)

A) Home computer users who want the most out of their machine do create liquid cooling systems for their PC, although they are rare. I bet tomshardware.com will give instructions on how to build/purchase one.

B) Few computers are practical without any power, in California or elsewhere.

C) I heard in the news that the X43-A was distroyed before it started the test, not after as it appeared you were saying. So far the rocket is the most likely reason it failed, not the aerodynamics of the test plane. Thus it appears that CFD may not have had any role in the X43-A failure.

D) I agree completely: we need computers that are affordable and fast. And we certainly all could use more experience and training!

E) Since both Intel and AMD expect 2 GHz machines by the end of this year, why do you feel they will stall out at 2 GHz? I think the momentum will push it to 4-5 GHz before we get any stalling (IE research these companies have been doing for the last few years will result in faster computers before any cut back on new research funding will have an impact). But that is just my humble opinion.

Either you completely missed my point, or you are being somewhat sarcastic ;-). If you could run a G4 at 1.3GHz or the P4 at 800 Mhz then you can get an idea of the advantage of the CPU design.

So, comparing different vendors processors at the same clock speed is not meaningless. I would concede that the comparison is somewhat academic, as it only gives you an idea of the performance of the overal chip design relative to another vendor, and your not going to purchas a P4 1.7GHz and run it at 1.3GHz right!

It is a well known fact that AMD will reduce their power consumption, by further reducing their die size. and making architectural changes. Something like this is not all that difficult for a decent CPU manufacturer. I can't remember the code name of the Athlon 4, but it's due out pretty soon. Simply check their roadmap at Toms hardware.

I agree that Jonas benchmarks a genuine, yet somewhat suprising. I'll be interested to run some things on a P4 once I have access to one.

Dan

(1). Faster computer will cost more. and I think, you will agree with me on that. (2). The reason why it is getting cheaper, is " most computers have been made in Taiwan" and that kept the cost low. (3). The transition is being made now to sustain that low cost idea: to make the PC in the third world countries, including China. (4). But with the difference in political systems (between US and the third world coumputer makers), someting is going to happen. (5). First possibility is: US will develop faster CPU, and the cost of computer will stay low by moving to the countries with cheaper labor cost. (not likely to happen,because US does not like the third world countries to have high tech CPU. (6). The other possibility is: slow down the development in CPU speed, and lower the cost of computer. (not likely to happen, becuase the companies in Taiwan are not going to make enough money to survive through selling hardware alone.) (7). In either way, the CPU companies will try to co-exist instead of pushing the clock speed to a higher level. Average business companies are not going to benefit from this CPU upgrade, because they need faster computer to do word processing. (8). And even if they can use it to speed up the network, the users will have to pay the price, thus higher cost for the customer. (9). It is like B-2 bomber case, they can continue to improve the aircraft, until the government can afford only one aircraft. The only reason why CPU maker can continue to push the clock speed higher in the past, is because makers in Taiwan are helping the world market to keep the cost low. Without that, CPU is not going to find a motherboard. (10). So, it is the end of the boom era for the CPU and the PC market. Unless, the cost is going to be absorbed by the user of PC, which will include the new CPU research cost, and the cost of living increase in Taiwan. It is just common sense.