I ran a simple test using Passmark Performance Test v 6.1. Just one video card in the system. This test anyone can do on their own system.

The top and middle test results used the x16 slot, the bottom bar is today's x4 slot performance. The top and bottom test are the same Gigabyte GTX 460. The 460's 2D performance is about the same in the x4 slot, except for GUI, while 3D - Complex performance (i.e. something like 'crunching') fell off quite a bit.

One unexpected thing noted before I started Passmark's test. I ran GPU-Z 0.5.4 and it said the 460 (in the x4 slot) is reporting it is capable of x16 v1.1 performance. I expected to see x16 v2.0 or 2.1 capability! I hope that's just something in GPU-Z and not really a 460 issue...



Lt

One unexpected thing noted before I started Passmark's test. I ran GPU-Z 0.5.4 and it said the 460 (in the x4 slot) is reporting it is capable of x16 v1.1 performance. I expected to see x16 v2.0 or 2.1 capability! I hope that's just something in GPU-Z and not really a 460 issue...

GPU-Z also reports my GTX460 as supporting PCI-Express x16 v1.1 (But it's running at x8 v1.1 as the HD5770 shares the lanes)

Claggy

Thanks Claggy! It's good to know at least one other person sees the same report.

I looked at my motherboard architecture again. The x4 slot on the P43 board is connected to the ICH10R ("Southbridge") while the x16 slot connects to the P43 ("Northbridge") chip.

Would the folks who are commenting about their experiences post the connect info for the pci-e slots in their system?

It could help explain the differences between experiences.


Lt

You need to remember that CUDA crunching is a completely different use of the GPU to gaming. In gaming the PCI-E bridge speed, and bandwidth capability, will be important to the game resolution, frame rate and many other aspects. So, the differences between x16, x8, x4 and x1 slots will determine the graphics performance as much as the NV card series.

Assuming the graphics card is CUDA capable I still contend that the PCI-E slot speed is not that important, certainly between x16 and x8. It possibly is not that important at x4 for the WU crunch time. The x1 slot will slow the WU crunching down.

There is not a great deal of data to transfer across the bridge slot for CUDA crunching, but in gaming it is very important.

Software that tests the PCI-E slot speed does not reflect the use of the GPU/PCI-E combination for project crunching.

So, from my understanding a x4 PCI-E slot will possibly slow the apparent output (RAC) of the project marginally (up to say 3%).

---

It's good to be back amongst friends and colleagues

You need to remember that CUDA crunching is a completely different use of the GPU to gaming. In gaming the PCI-E bridge speed, and bandwidth capability, will be important to the game resolution, frame rate and many other aspects. So, the differences between x16, x8, x4 and x1 slots will determine the graphics performance as much as the NV card series.

The x16/x4 comparison just showed me that the 460 could not get the data fast enough in the x4 slot to perform anywhere near x16 capability.

BTW, I asked Passmark in Aug, 2009 if they were considering development of a CUDA/OpenCL test. Part of the reply said this "But it should be noted that the use of shaders (code which is compile and executed on the GPU) in the existing 3D test make use of the GPU's processing power."

Assuming the graphics card is CUDA capable I still contend that the PCI-E slot speed is not that important, certainly between x16 and x8. It possibly is not that important at x4 for the WU crunch time. The x1 slot will slow the WU crunching down.

Anyone know when pci-e 3.0 will start to appear in consumer pc products?

There is not a great deal of data to transfer across the bridge slot for CUDA crunching, but in gaming it is very important.

Software that tests the PCI-E slot speed does not reflect the use of the GPU/PCI-E combination for project crunching.

So, from my understanding a x4 PCI-E slot will possibly slow the apparent output (RAC) of the project marginally (up to say 3%).

I'm crunching now with the 460 in the x4 slot. Nvidia AP and MB wu's. Tomorrow afternoon will be 24 hrs. I'll see if I can detect any differences in real-world times then. I expect I won't be able to tell any difference in MB crunching. MB's just don't push the card, or i/o, hard enough. There may be noticeable differences in CUDA AP crunching, though.

Lt

Doubtful.

---

Doubtful.

Yep, very. :) Only two AP's completed since yesterday. One found 30 pulses and ended too soon. So, I don't see anything obviously different after 24 hrs.

I've never tried to do the repetitive testing that Sutaru and others do using the same mix of workunits in multiple configs. That may be the only way to see if there is any difference at all.

Lt

its the same thing as back when agp was still around. agp 4x was only slightly faster then agp 2x. and agp 8x was basically the same performance as agp 4x.