Dear freesurfer team
I'm currently investigating how to tweak 2 2 year old Dell 7400 Xeon machine's to run a large batch of fs jobs. Upgrading to Nehalem won't work, since it will require a new motherboard as well. However, we can replace our old non-fermi quadro card with a C2050.
As an alternative, I was thinking about buying two GTX590 (two times 580 and 1.5 GB ram) or two palit GTX580 (1 times 580 with 3GB ram) and running them in SLI. This will save quite some money, and intuitively, it feels that more separate cards will work better if you start batching more recon scripts at the same time (1 per CPU). You'll end up with 6 gb of vram in both cases compared to the 3 for the C2050, with 4 processors for the 590 and 2 for the 580. As I don't understand cuda and sli well, I may very well be completely wrong. I'm not sure fi that running jobs under sli will work well. Do you have any ideas on this solution compared to a single c2050?
thank you very much in advance,
Andries van der Leij
In my experience you'll have the benefit of using two cards simultaneously. That means you can run two recon-all with gpu acceleration simultaneously. The SLI is more important for graphics (cad, gaming) than for CUDA. So, I'd go for 2 GTX580
PPJ
--------------------------------------------------------------------- Pedro Paulo de Magalhães Oliveira Junior -- check our new iOS app: Cargohttp://itunes.apple.com/app/cargo/id444529879?ls=1&mt=8
On Fri, Aug 5, 2011 at 14:28, Andries R. Van Der Leij < a.r.vanderleij@gmail.com> wrote:
Dear freesurfer team
I'm currently investigating how to tweak 2 2 year old Dell 7400 Xeon machine's to run a large batch of fs jobs. Upgrading to Nehalem won't work, since it will require a new motherboard as well. However, we can replace our old non-fermi quadro card with a C2050.
As an alternative, I was thinking about buying two GTX590 (two times 580 and 1.5 GB ram) or two palit GTX580 (1 times 580 with 3GB ram) and running them in SLI. This will save quite some money, and intuitively, it feels that more separate cards will work better if you start batching more recon scripts at the same time (1 per CPU). You'll end up with 6 gb of vram in both cases compared to the 3 for the C2050, with 4 processors for the 590 and 2 for the 580. As I don't understand cuda and sli well, I may very well be completely wrong. I'm not sure fi that running jobs under sli will work well. Do you have any ideas on this solution compared to a single c2050?
thank you very much in advance,
Andries van der Leij
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Partners Compliance HelpLine at http://www.partners.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail.
2011/8/5 Pedro Paulo de Magalhães Oliveira Junior ppj@netfilter.com.br:
In my experience you'll have the benefit of using two cards simultaneously. That means you can run two recon-all with gpu acceleration simultaneously. The SLI is more important for graphics (cad, gaming) than for CUDA. So, I'd go for 2 GTX580
CUDA doesn't use SLI at all - you don't want it. So long as each card has enough RAM to handle one recon job, having more cards and using the environment variable to select the GPU is certainly the way to go. I don't think that Freesurfer is particularly heavy on the PCIe bus, so you should even be able to get one of the dual GPU cards and get good performance. However, Nehalem does help the data reorganisation required for the GPU.
Richard
Hi pedro and Richard,
Thank you for the responses! The T7400 has just one gpu slot, so I'll probably just buy the single 590, which itself is a dual card. The 2050 doesn't seem to be worth it on these older systems (ddr2, etc).
For my supervisor, what kind of performance increase should I expect if we buy a newer nehalem system?
Regards,
Andries van der leij
Sent from my iPhone
On Aug 5, 2011, at 20:21, R Edgar freesurfer.rge@gmail.com wrote:
2011/8/5 Pedro Paulo de Magalhães Oliveira Junior <ppj@netfilter.com.b r>:
In my experience you'll have the benefit of using two cards simultaneously. That means you can run two recon-all with gpu acceleration simultaneously. The SLI is more important for graphics (cad, gaming) than for CUDA. So, I'd go for 2 GTX580
CUDA doesn't use SLI at all - you don't want it. So long as each card has enough RAM to handle one recon job, having more cards and using the environment variable to select the GPU is certainly the way to go. I don't think that Freesurfer is particularly heavy on the PCIe bus, so you should even be able to get one of the dual GPU cards and get good performance. However, Nehalem does help the data reorganisation required for the GPU.
Richard
I would go for a complete new PC. You won't benefit much of a GTX590 in an old PC
PPJ
--------------------------------------------------------------------- Pedro Paulo de Magalhães Oliveira Junior -- check our new iOS app: Cargohttp://itunes.apple.com/app/cargo/id444529879?ls=1&mt=8
2011/8/5 Andries R. Van Der Leij a.r.vanderleij@gmail.com
Hi pedro and Richard,
Thank you for the responses! The T7400 has just one gpu slot, so I'll probably just buy the single 590, which itself is a dual card. The 2050 doesn't seem to be worth it on these older systems (ddr2, etc).
For my supervisor, what kind of performance increase should I expect if we buy a newer nehalem system?
Regards,
Andries van der leij
Sent from my iPhone
On Aug 5, 2011, at 20:21, R Edgar freesurfer.rge@gmail.com wrote:
2011/8/5 Pedro Paulo de Magalhães Oliveira Junior <ppj@netfilter.com.b r>:
In my experience you'll have the benefit of using two cards simultaneously. That means you can run two recon-all with gpu acceleration simultaneously. The SLI is more important for graphics (cad, gaming) than for CUDA. So, I'd go for 2 GTX580
CUDA doesn't use SLI at all - you don't want it. So long as each card has enough RAM to handle one recon job, having more cards and using the environment variable to select the GPU is certainly the way to go. I don't think that Freesurfer is particularly heavy on the PCIe bus, so you should even be able to get one of the dual GPU cards and get good performance. However, Nehalem does help the data reorganisation required for the GPU.
Richard
2011/8/5 Andries R. Van Der Leij a.r.vanderleij@gmail.com:
For my supervisor, what kind of performance increase should I expect if we buy a newer nehalem system?
Freesurfer as a whole is quite keen on the improved memory subsystem offered by Nehalem. A figure of 2x (CPU only stream) is floating in my mind, but I don't recall the exact two setups.
Richard
freesurfer@nmr.mgh.harvard.edu
-
Andries R. Van Der Leij -
Pedro Paulo de Magalhães Oliveira Junior -
R Edgar