Influence of PCI Express Lanes to GPU Computing

[Brain]

Hello,

does anybody have experiences with the influence of PCI Express lane count to GPU Computing? Especially with regard to CUDALucas but mfaktc too. All I found is this:

Originally Posted by axn

Improved memory should have a more pronounced impact on CUDALucas.

Originally Posted by BigBrother

Well, The Card is now inserted into a PCI-E 2.0 x16 slot, and my brain surgery skills allowed me to fix a bent pin on the CPU socket so my memory is back at dual channel again.

One instance of mfaktc is now taking +-70% GPU instead of the 74% I reported yesterday, and nVidia's Visual Profiler shows transfer rates of 6Gb/s instead of 3 Gb/s, but since the amount of data to transfer is relatively small, there's no earth-shattering improvement. I could run the same benchmark I did yesterday again if James would like me to do that.

I have to decide whether I put my second GPU next to the first (X8 = 8 Lanes but closer and hotter) or to the boards bottom (X4).

I could simply try but I would like not to reinvent the wheel. I'd love to use the X4 slot for thermal reasons.

My guess: CUDALucas is not very intensive regarding CPU to GPU transfers - mfaktc will use a bit more. Any concrete numbers would be brilliant.

Thanks, Brain

[TheJudger]

For mfaktc you've to transfer approximately four bytes per FC (factor candidate). So when you're GPU is good for 250M FCs per second than you'll need 1GB/s of PCIe bandwidth. Ofcourse there is some overhead but I guess it is save to ignore the overhead.

Oliver

[Brain]

Originally Posted by TheJudger

For mfaktc you've to transfer approximately four bytes per FC (factor candidate). So when you're GPU is good for 250M FCs per second than you'll need 1GB/s of PCIe bandwidth. Ofcourse there is some overhead but I guess it is save to ignore the overhead.

Oliver

Me happy to have PCI Express v3.0:
http://en.wikipedia.org/wiki/PCI_Express
Capacity Per lane (each direction):

• v1.x: 250 MB/s (2.5 GT/s)
• v2.x: 500 MB/s (5 GT/s)
• v3.0: 1 GB/s (8 GT/s)
• v4.0: 2 GB/s (16 GT/s)

16 lane slot (each direction):

• v1.x: 4 GB/s (40 GT/s)
• v2.x: 8 GB/s (80 GT/s)
• v3.0: 16 GB/s (128 GT/s)

[henryzz]

Makes me wonder what performance I would get if I put a pcie 2 card in my pcie1.1 slot. I discovered a while back that they work just at the slower speed.

[Stef42]

I found that running mfaktc (GTX460) in PCI-E x1 mode made the performance about 1/2 of that in a PCI-e x16 slot.
Using the same card in x4 mode gave roughly the same performance as on x16.

[LaurV]

For me no difference on x8/x16 (PCIe 2.0). Have no x4 cards/slots.

[henryzz]

My 1.1 x16 slot should theoretically run at the same speed as a 2.0 x8 slot. Sounds like I have more than enough bandwidth for a GTX460.