Sign in with
Sign up | Sign in
Your question

I7-950 vs Q9550 for CFD calculations

Last response: in CPUs
Share
December 13, 2009 5:36:52 AM


Hi,

I have a Q9550 CPU right now and I am using all of the four CPUs as a means for Computation purposes of CFD.

Every case I run, takes nearly a complete week! So I was wondering if buying an i7-950 would solve my problem.

How much faster is an i7-950 than a Q9550 for multi-core calculations ?

Thanks in advance.
December 13, 2009 6:09:39 AM


It is very important for me. Does'nt anyone has a suggestion?
m
0
l
a b à CPUs
December 13, 2009 9:22:33 AM

It would be a heck of an improvement, and if you're willing to shell out the cash, it would definitely give you a noticeable speed boost.
m
0
l
Related resources
a b à CPUs
December 13, 2009 10:38:20 AM

depends if the cpu is the only bottlekneck - if your using XP thats the first thing to dump and then fill the system with 4+ gb of ram, if its hdd related an SSD would help but if its totally cpu then yes the i7 would blitz the C2Q, and overshocked - generally overclocking + servers, business and reliability requirements dont mix
m
0
l
December 13, 2009 10:56:05 AM

Why the i7-950 (not the i7 920)? it is much cheaper and it you plan to overclock then there is not much difference between them.

If a stock q9550 takes a week to complete a task then a stock i7 950 may take about 4.5-5 days. A 6 core westmere or 2 xeon processors will be faster.
m
0
l
December 13, 2009 2:25:31 PM


Hi,

Thanks for the replies.

I am using Linux Fedora Core 9 for my calculations. Is it possible to overclock there to? If yes, how much time can the CPU be put to overclock? (As far as I know CPU can be used continuously, in normal mode)

Thanks in advance
m
0
l
a c 104 à CPUs
December 13, 2009 4:10:13 PM

+1 with basket687, i7-920 would make a more sensible purchase at a good price with plenty of headroom for overclocking to crunch its way through calculations.
m
0
l
a c 113 à CPUs
December 13, 2009 6:57:33 PM

mmah said:
Hi,

I have a Q9550 CPU right now and I am using all of the four CPUs as a means for Computation purposes of CFD.

Every case I run, takes nearly a complete week! So I was wondering if buying an i7-950 would solve my problem.

How much faster is an i7-950 than a Q9550 for multi-core calculations ?

Thanks in advance.
You already received several suggestions, but nobody asked you what your goal is. The task that currently completes in almost one week should be completed in how many days? Is the CPU the bottleneck?
m
0
l
December 13, 2009 7:10:34 PM


My goal is to improve the speed of the calculations by an amount of like 2 times.

And I think that yes the CPU is the bottleneck, because I have 8 GB of RAM (DDR3), of which only 50 percent is used during work.

So do you think running on an overclocked i7 processor for 4 or 5 days is possible?
m
0
l
a c 79 à CPUs
December 13, 2009 7:43:08 PM

cuda? or directcompute?
m
0
l
December 13, 2009 7:51:22 PM


If cuda means using the graphic card gpu for calculations, I am going to have to say it is direct compute.
m
0
l
a c 113 à CPUs
December 13, 2009 7:58:42 PM

The answer definitely is no. You might gain 1 or 2 days, but it won't run in 1/2 the time.
m
0
l
December 13, 2009 8:01:10 PM


Is the gain with overclocking or without it? For how much time can i overclock?
m
0
l
December 13, 2009 8:05:40 PM


2 Questions then:

1) For ordinary use, is the gain with overclocking or without it? For how much time can i overclock?

2) My Graphic card is NVIDIA GTX260. Can I use it for direct compute?

m
0
l
a c 113 à CPUs
December 13, 2009 8:09:10 PM

I based my comments on a non-overclocked CPU, but both CPUs could be overclocked. On the other hand, you probably can't overclock as much as some people do because your system has to be very stable. Having a faster system that crashes after 3 days doesn't really help, does it?

Edit: If it mostly uses the ALU, then the performance gains could reach 40-45%. If you look at Tom's CPU charts, the Q9550's performance index is 76% of an i7-950. But it only is 57% on the processor arithmetic benchmark. I don't know how that translates to your application.
m
0
l
December 13, 2009 8:28:49 PM


Thanks very much for informative reply GhislainG. Actually the solver I am using "PBiCg" (Krylov subspace solver!) with DILU preconditioner. It is one of the solvers of the OpenFOAM code. And I would be thankful if you could post the link of the article you mentioned.
m
0
l
a c 79 à CPUs
December 13, 2009 8:32:06 PM

latest nvidia drivers support direct compute, not sure what OS you need though, win 7 definately ok. vista not sure... i'm out of my depth now... :) 
m
0
l
a b à CPUs
December 13, 2009 9:27:54 PM

13thmonkey said:
latest nvidia drivers support direct compute, not sure what OS you need though, win 7 definately ok. vista not sure... i'm out of my depth now... :) 

That doesn't mean that every app can use it, and I would go as far as to say that in this case, it probably can't be used.

Oh, and I probably wouldn't go for too much of an overclock. With good cooling, you could push it up a bit, but since you'll be stressing it for a week at a time (or at least several days), you'll need absolutely rock solid stability. You can definitely get an overclock stable enough, but it would take some tweaking and testing and quite a bit of time running stability tests.

I would say that you could realistically expect a gain of 20-50%, depending on what the application is currently bottlenecked by (FPU, ALU, memory bandwidth, etc).
m
0
l
a b à CPUs
December 13, 2009 9:49:53 PM

IF you are TRULY building a new PC for CFD I highly recommend you consider getting a 2P Nehalem based PC.

@OP: As far as GPU goes, I doubt any OCing,etc done to the GPU will help.

Personal note from me to OP:
I use SolidWork's COSMOS suite along with Thermal Desktop,etc. I built a mid high end CAD/CFD workstation for a neighbor based on 2P Nehalem (see: http://www.tomshardware.com/forum/267080-31-build ) comparing the 2P system to my OCed i7 system( see sig), it runs the CFD apps about 15-25 (or ~32 in the case of FloWorks)% faster.

@OP:
List budget for a new system (CPU + RAM + Motherboard + may be PSU; you can probably re use case, HDDs,etc).
m
0
l
December 14, 2009 12:27:48 AM

Well, I'm not the one writing the CFD application but I do manage a cluster that one type of the applications ppl run on it is CFD. Anyway, if your application with your current input runs at one week now on a Q9550, the best you can do with an i7-950 is half that time, taking into consideration that you have modified it to take advantage of hyperthreading. A typical CFD application can be optimize to run well on hyperthreading. If not you'd gain around 2 days at best.

Other things you might wanna consider are:
- How's the RAM usage, it should be quite easy to check with the top application on linux or task manager on Windows. If your application is starve of RAM, it will run much slower.
- Move your application to a cluster if possible. Or build a small cluster with around 4 to 8 node consist of DIY workstations.
- Move your application to CUDA: it does support C and Fortran, and it has a full toolset including profiler application. I would not recommend OpenCL at the moment, its toolset is not up on par with CUDA as of now and if you use Fortran, OpenCL is not for you.
Both of the latter two require rewriting of your appliaction, with the last one require major overhaul and an understanding of GPGPU programming to have the best possible result. You should also consider getting a GTX280 as the weaker has less than phenomenon double precision performance.
m
0
l
a b à CPUs
December 14, 2009 2:43:55 AM

Shadow703793 said:
IF you are TRULY building a new PC for CFD I highly recommend you consider getting a 2P Nehalem based PC.

@OP: As far as GPU goes, I doubt any OCing,etc done to the GPU will help.

Personal note from me to OP:
I use SolidWork's COSMOS suite along with Thermal Desktop,etc. I built a mid high end CAD/CFD workstation for a neighbor based on 2P Nehalem (see: http://www.tomshardware.com/forum/267080-31-build ) comparing the 2P system to my OCed i7 system( see sig), it runs the CFD apps about 15-25 (or ~32 in the case of FloWorks)% faster.

@OP:
List budget for a new system (CPU + RAM + Motherboard + may be PSU; you can probably re use case, HDDs,etc).


That's a good point actually. The OP could definitely take advantage of a 2P system, so it might be the way to go.
m
0
l
December 17, 2009 11:34:36 AM


Hi again,

Is hyper-threading to 8 threads possible with an i7-950 in Linux as well as Windows? Doesn't it require a specific program?

Thanks in advance for the replys

m
0
l
a b à CPUs
December 17, 2009 11:43:32 AM

It's a feature of the CPU itself - it should be completely OS independent.
m
0
l
!