|
[Sponsors] |
May 22, 2001, 17:04 |
Best CFD machine:Athlon or P4?
|
#1 |
Guest
Posts: n/a
|
I want to buy the fastest number crunching PC:
1)1.33GHz AMD Thunderbird (266 Mhz FSB) with 512MB PC2100 DDR memory 2)1.3GHz AMD Thunderbird with 512MB PC133 133MHz memory 3)1. 3GHz Pentium 4 with 512MB 800MHz Rambus All systems will have a 7200RPM 30GB Ultra/100 hard drive and Win 2000. Which system is the best for computational fluid mechanics applications? |
|
May 22, 2001, 18:35 |
Re: Best CFD machine:Athlon or P4?
|
#2 |
Guest
Posts: n/a
|
It depends on the algorithm that your CFD solver is using. A few programs might be limited by raw CPU speed (which is not the same as clock speed), but, in my experience, memory size, bandwidth and latency issues are often important considerations in this type of evaluation. Really, you should benchmark these computers with the actual CFD code that you plan to run and with some typical problems that you want to solve. Hypothetically, for example, you could see if the typical problems that you're solving that might barely fit onto a 1 GB SDRAM machine might not necessarily fit onto an RDRAM machine, since the same money would purchase less memory, and cause your code performance to degrade considerably as virtual memory on disk is used. You haven't mentioned whether cost is a consideration. A top-of-the-line Alpha RISC machine will beat any <tt>x86</tt> machine in a pure performance test of a single processor. But top of the line performance costs top of the line dollars and the relationship isn't linear, either OTOH, if your application parallelizes, then you might get a more favorable price/performance from buying several cheap <tt>x86</tt> machines and making a Beowulf cluster of them. This site shows some interesting comparisons between the x86 platforms that you're considering. They include results for both Athlon and P4 for a few of the LINPACK kernel algorithms which may constitute the bulk of the work done by your CFD code (again, depending on the algorithm that your code is using). BTW, if you do get some performance numbers for these machines with a particular code and a precisely-defined benchmark problem, then there are probably a number of readers here that would be interested in seeing those results! |
|
May 22, 2001, 18:46 |
Re: Best CFD machine:Athlon or P4?
|
#3 |
Guest
Posts: n/a
|
Thanks for the quick response. Money is an issue but the time to get a converged solution is more important. I don't think that we can afford the Alpha Risc machines if these machines are the Sun, IBM, SGI workstations. We are willing to pay for Rambus if its results in faster solution than 266/200Mhz or PC133 memory. I wish we could compare both machines but we will buy these machines from a local vendor who will build them when we order so there is no opportunity to test unless I can convince my boss to buy one of each but then we have two diffirent machine whose parts are not interchangeable. We intend to parallelize in the future.
|
|
May 22, 2001, 19:10 |
Re: Best CFD machine:Athlon or P4?
|
#4 |
Guest
Posts: n/a
|
(1) will beat (2) for sure
|
|
May 22, 2001, 19:56 |
Re: Best CFD machine:Athlon or P4?
|
#5 |
Guest
Posts: n/a
|
I think there was a long thread on this subject just a week or so ago.
Have you checked the discussions in the archieves and near the bottom of the 'current 500' posts? |
|
May 23, 2001, 01:43 |
Re: Best CFD machine:Athlon or P4?
|
#6 |
Guest
Posts: n/a
|
(1). As I said before, PC is not designed for CFD at all. (2). If you are writeing codes, then anyone of the machine is good enough. The cheaper the better. It is more than adequate for cfd work. Unless you are dealing with reacting flows, or complex multi-stage machines (even these problems can be simplified) (3). If you are using commercial codes or someones' codes then, it is a good idea to get the feed back from the author of the codes. Sometimes, these authors have special considerations about the hardware usage and the operting system. (4). You also can look into the problem itself and see whether there are different ways to write the program to speed up the convergence. (this is always problem dependent)
|
|
May 23, 2001, 02:46 |
Re: Best CFD machine:Athlon or P4?
|
#7 |
Guest
Posts: n/a
|
I recently got benchmarks comparing a 1.5 GHz P4, an 1.33 GHz Athlon with PC133 SDRAM and a 1.33 GHz Athlon with DDR SDRAM (PC2100). These benchmarks were made with Fluent - the results might be different with another code. Here are the numbers:
1.5 GHz P4, RDRAM: 1.33 1.33 GHz Athlon, DDR: 1.05 1.33 GHz Athlon, PC133: 1.00 As you can see the P4 is 30% faster than then Athlon. This is very supricing to me. Note also that the code used (Fluent) is not at all optimized for the P4. The benchmarks were run on RedHat Linux. As a side note - the 1.7GHz P4 is probably the fastest single-CPU machine available now, beating the most expensive workstations from HP, Compaq and SGI! |
|
May 23, 2001, 04:16 |
Re: Best CFD machine:Athlon or P4?
|
#8 |
Guest
Posts: n/a
|
Most of the CFD codes I've come across seem to pretty much follow specint single CPU benchmarks, which are regularly updated at:
http://www.ideasinternational.com/be...int_s2000.html (Last time I checked out HP's web site there was no spec or price for the new PA-8700 powered workstations though.) - Steve |
|
May 23, 2001, 07:38 |
Re: Best CFD machine:Athlon or P4?
|
#9 |
Guest
Posts: n/a
|
I have an article from a computer magazine where they benchmark a P4 at 1.5GHz against an Athlon at 1.2GHz. These are not CFD tests but they can be helpful.
The first test is the kernel compilation test, and in this case, the Athlon was 1.5 times faster than the P4. The second test is for MP2 compression. The timings were not given. But they said that the P4 was far more superior than the Athlon. The reason why the P4 was superior was given and is directly related to CFD. The P4 uses a trace back L1 chache which is very usefull for loops (dot produts....), the instructions are only decoded once and not everytime the loop is used. Which speed up the code when many little loops have to be proceeded. Regards |
|
May 23, 2001, 09:31 |
Re: Best CFD machine:Athlon or P4?
|
#10 |
Guest
Posts: n/a
|
Steve,
I have a pair of Athalon machines in my office that I am very happy with. Both have 768Mb of RAM, and I would recommend getting more memory than you think you need. Nothing will spoil your day faster than running out of memory and having to run out of core (or having the code crash). In this company of ~150 people we have over 250 computers. We have at least 50 Athalons (there's a cluster of 32 of them in one room). On the other hand, we have 1 Pentium 4, and I've heard (second hand) that it is fairly unstable. Regards, Alton |
|
May 23, 2001, 10:10 |
Re: Best CFD machine:Athlon or P4?
|
#11 |
Guest
Posts: n/a
|
The results are somewhat disappointing, I would say. If you look at LINPACK results the P4 should be a lot better due to the high memory bandwith. We ran a benchmark as well, using our own software, and the result was even a little worse from what you show here. I don't think it is the CPU, which is to blame, though. There is a huge potential for optimization in the codes, or the compilers, I guess. From a pragmatic point of view the Athlon seems to be the choice, because it is a LOT cheaper. At least that is what we decided for ourselves.
|
|
May 23, 2001, 11:09 |
Re: Best CFD machine:Athlon or P4?
|
#12 |
Guest
Posts: n/a
|
Due to huge pipeline length of the P4, low clock speed P4 machines cannot compete very well on a clock per clock basis. That will change as the P4 goes past the 2 GHz mark. Thus, I would never recommend that 1.3 GHz machine.
However, on June 4th, 2001 you can buy a dual processor P4 at 1.7 GHz which should give you much better performance than any other non-Alpha computer for CFD work. A price example at Dell: Dual 1.7 GHz, 80 GB hard drive, 128 MB RAM, no monitor for $3894. Then throw away the memory Dell sold and add in 1GB RAM bought online for $796 gives a total of $4690 plus shipping of course. If money is a big concern, dual Athlons will also be available by the end of summer (the memory will cost $400 less and the processors will each be $200 less roughly). |
|
May 23, 2001, 11:14 |
Re: Best CFD machine:Athlon or P4?
|
#13 |
Guest
Posts: n/a
|
The DDR memory should be twice as fast as the PC133 memory. Unfortunately, real world performance shows much less improvement (I've seen 10%-30% quotes). RDRAM is the fastest available (currently P4 computers require it, and no other system uses it), but I don't think memory speed is the current bottleneck in many CFD commercial programs and RDRAM is much more expensive.
Try your best to convince your boss to try both machines, you could even try renting one for a month and do some tests (about $100). |
|
May 23, 2001, 16:41 |
Re: Best CFD machine:Athlon or P4?
|
#14 |
Guest
Posts: n/a
|
I think Dr. Strangelove's advice is good, in the sense that different codes seem to suit different hardware configurations. I'm also agonising over this choice at the moment, and it has been highly educational. Some conclusions so far:
1. Beware of "synthetic" benchmarks. For example, the standard linpack SGEFA/SGESL DGEFA/DGESL benchmark code that has been used to generate some valuable information for Anandtech's web site in particular, is very valuable for gaining an understanding of the impact of FPU, cache and memory bandwidth. However, it doesn't tell you all that much about CFD performance! One would think that a CFD-code would operate on the right hand side of the graph, where performance is totally determined by memory bandwidth. This would give a 2:1 performance advantage to the P4, but in fact, for CFD work the advantage of a 1.7 P4 over a 1.33 Athlon appears to be somewhere between 1.1 to 1.3 : 1 When you think about it, it makes sense - the Linpack program uses Gauss elimination to solve a large set of linear equations, and all CFD codes use some kind of iterative scheme, which is far less memory intensive. 2. The BS3DVW_65 scores are also somewhat "misleading". Although more representative of a CFD code, it still seems to result in an artificial advantage for the P4 (about 1.5 : 1). My conclusion is that it favours a fast ot large L2 cache (P4 has a faster L2 cache), based also on the fact that my old P-II (512 kB L2) is faster than several P-III's with 256 kB L2! 3. Benchmarking CFD-Fastran on a 1.5 GHz P4, and extrapolating results from a 1100 MHz Athlon to a 1333 Thunderbird with DDR, indicates that performance should be quite similar. But the Athlon with DDR is significantly cheaper than a P4 with RDRAM. So, the choice for a single CPU machine or a cluster comes down to the AMD. 4. However. Software licensing for a commercial code is MUCH more expensive than the hardware costs that we are talking about here, so you really do need to take that into account when trying to get to the optimum setup. (For example, CFDRC's parallel licensing is relatively attractive, compared to some other codes, and their single node-locked license heavily favours a dual-CPU system) |
|
May 23, 2001, 17:34 |
Re: Best CFD machine:Athlon or P4?
|
#15 |
Guest
Posts: n/a
|
Charles Crosby: thank you for some great advice. The CFD code that I will be using, WIND, is free.
|
|
May 23, 2001, 17:36 |
Re: Best CFD machine:Athlon or P4?
|
#16 |
Guest
Posts: n/a
|
Jim Park: I will checkout the older posts or archive. Thanks.
|
|
May 23, 2001, 17:38 |
Re: Best CFD machine:Athlon or P4?
|
#17 |
Guest
Posts: n/a
|
John Chien: The CFD code that I use, NPARC and WIND were written for the Cray and Unix workstations. The developers have little knowldege of PC technology.
|
|
May 23, 2001, 17:40 |
Re: Best CFD machine:Athlon or P4?
|
#18 |
Guest
Posts: n/a
|
Jonas Larson: Thanks for sharing your benchmark.
|
|
May 23, 2001, 17:47 |
Re: Best CFD machine:Athlon or P4?
|
#19 |
Guest
Posts: n/a
|
Steve Amphlett: This comparison table is interesting but confusing since I see no differences between systems 8 & 13 or systems 14 & 17, yet they have different performances. It may be that systems specs are not completely given.
|
|
May 23, 2001, 17:52 |
Re: Best CFD machine:Athlon or P4?
|
#20 |
Guest
Posts: n/a
|
Alton: I am partial to the Athlon since I have a 1.2 GHz system with 768MB as my personal system. It works great. But my boss is mesmerized by the 800MHz Rambus.
|
|
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
CFD Wiki - We Need More Help! | Jonas Larsson | Main CFD Forum | 0 | September 26, 2005 08:11 |
Where do we go from here? CFD in 2001 | John C. Chien | Main CFD Forum | 36 | January 24, 2001 21:10 |
ASME CFD Symposium, Atlanta, July 2001 | Chris R. Kleijn | Main CFD Forum | 0 | August 21, 2000 04:49 |
Since Last June | John C. Chien | Main CFD Forum | 3 | July 12, 1999 09:38 |
Which is better to develop in-house CFD code or to buy a available CFD package. | Tareq Al-shaalan | Main CFD Forum | 10 | June 12, 1999 23:27 |