Folding with a GPU in a PCI-e 16x @ PCI-e 4x

[Lowrez]

Would there be a performance hit? how much of one?

There are some motherboards out that have 4 PCI-e 16x slots but some of the slots run at 4x. I was thinking about getting one and putting 4 8800gt's in it for folding.

Also I want to get an Intel board instead of an Nvidia one. Do you think there will be driver issues to get all 4 cards to work?

Potentially it could be a monster folding box. 20,000ppd if all went well for under 2000$ US(maybe even lower $$)

What do you think?

[Marvin_The_Martian]

Even an 8x slot would slow the client, been posted on the offical forum. You can test for yourself, just reboot with your card in x4 or x8 mode set in bios and compare ppd. The gpu/cpu communications are even what seem to hinder anything above 9800gtx+ as they don't get more work done in comparison.

Pretty late edit: I misread the entire thing as what was said was the exact opposite, you could do atm with a 1x slot for bandwidth. Though it might change if for instance Ati get's all 800 sp's working and you need to transfer allot more data.

[Lowrez]

Well after some checking around I have found that you can build a 780i board with 3 PCIe 16x slots and 3 9600gso's and a Q6600 in a rig for between 1000 to 1200 dollars and get (being conservative) 15,000ppd. Thats pretty damn good.

[Marvin_The_Martian]

I need to correct myself, I think a moderator on the folding forums on behalf of somene from Nvidia has claimed even an 4x slot shouldn't interfear with the fah data. First I thought it be impossible, but thinking about it, the amount or ram used by the gpu client observed in taskman never goes really high. Maybe when the work units increase in size a 4x slot would result in slowdowns though idk. I'll try and link you to the comment in a bit, have to finish my coffee before I'm really awake :)

[mstp2009]

Current NVGPU2 WUs only use 6MB (yes, 6) of the video card RAM, and all the computations stay on-chip. There is virtually no CPU-GPU communication that is necessary.

In theory, with the current small WUs, even 1X PCI-E has ample bandwidth.

This all might change w/ larger WUs, especially if they get larger than a card's onboard RAM, but for now it is not an issue.

[Stijndp]

Current NVGPU2 WUs only use 6MB (yes, 6) of the video card RAM, and all the computations stay on-chip. There is virtually no CPU-GPU communication that is necessary.

In theory, with the current small WUs, even 1X PCI-E has ample bandwidth.

This all might change w/ larger WUs, especially if they get larger than a card's onboard RAM, but for now it is not an issue.

What you say doesn't make much sense tbh :) They aren't going to let people download wu's of +512MB every 4hours. Maybe max 20MB or 30MB for longer wu's.

[mstp2009]

What you say doesn't make much sense tbh :) They aren't going to let people download wu's of +512MB every 4hours. Maybe max 20MB or 30MB for longer wu's.

You don't DL a 1/2GB WU every 4h. :rofl::ROTF:

Have you SEEN how compressed the large WUs for SMP are? It's crazy compression



And to answer the OP's original question - - I now have Dual 8800GT 512MB cards folding in a P35, one on a 16X and one on a 4X slot. The PPD from the cards is IDENTICAL.

[Lowrez]

Hey mstp2009 thanks for all the info. You have been a big help. Especially the link to your forums where the dual gpu hack was first shown.

So now the race begins to see who can make the most ppd outta one mobo. Just think if you put 4 9800gx2's in one box (how would you cool that?). Now that would be something.

Diseases better watch out!

[Marvin_The_Martian]

Current NVGPU2 WUs only use 6MB (yes, 6) of the video card RAM, and all the computations stay on-chip. There is virtually no CPU-GPU communication that is necessary.

In theory, with the current small WUs, even 1X PCI-E has ample bandwidth.

This all might change w/ larger WUs, especially if they get larger than a card's onboard RAM, but for now it is not an issue.

Thanks for the info I was a litle slugish and should have posted that info when I finished my coffee :coffee:

[mstp2009]

Thanks for the info I was a litle slugish and should have posted that info when I finished my coffee :coffee:

Stanford has leaked a little more info on this in the folding forum.

They are going to specifically design WUs that make use of the "majority of cards out there" and do not intend to push RAM limits.

We don't know if this means 256MB cards or 512MB cards will be the design base, but it is good for us folders b/c this pretty much sums up that PCI-E bandwidth will never be a limiting factor.

[Marvin_The_Martian]

I've got a new account, but I'm not a stranger there :up:

Been folding for quite some time now.

[mstp2009]

I've got a new account, but I'm not a stranger there :up:

Been folding for quite some time now.

Wasn't implied as you were. In fact, the post was directed at the audience in general, and not you specifically.

Was simply posting more info about GPU RAM requirements for the cards so everyone knew.

[Marvin_The_Martian]

I took it as directed to me specificly since you quoted me and wth are we arguing about anyway :confused:

I think it will be 512, as when they talked about the x1900 series they mentioned the performance of the 256mb version as surprisingly addequate. Meaning that they were initially thinking about the 512mb as being the lower limit, maybe because that's how much ram they'd need for the number of atoms and itterations involved with the clients they had before ( single core and smp ), and I still believe the amount of memory needed isn't really architectural dependant but rather dependant on the subject being simulated. Idk if it makes allot of sence but it sounds plausible to me?