External Video cards and Performance

[SunDevilDave]

I would like to know why an external graphics card would be slower in performance than an internal card.

At work we have CUBIX boxes. Basically they allow you to upgrade your video cards without replacing your power supply; they also allow the placement of mutliple graphics cards that physically would not fit in a desktop computer. The CUBIX boxes join to the motherboard through PCI-e x16 extension cards and molex cables. The molex is capped at what I believe is 5.0 Gbps. We have two connections to the host machine and each CUBIX box has two 580 GTX cards....so 4 Cuda devices at render; 2048 cuda cores.

Bart Gawboy came out to our facility last week to teach us more about MR in general. He noted that compared to the Maximus system, our system seemed slower on comparable scenes using Iray. Maximus, using both the Tesla and Quadro has 896 Cuda Cores.

So where is the difference coming from? Here are my options:

1. Hardware limited somehow by the molex connection.
2. 4 Cudas devices (one of which is used for display). Is there a significant performance hit if the GUI display is dragging down the whole render?
3. Nvidia handicapping the drivers. Is the difference because the FP64 is being capped?
4. Is Maximus built on PCIe 2.0 or 3.0? and is the max bandwidth of the PCIe being used by the system?

Answering just one of those questions would be helpful.

[Phillip Miller]

In theory, there should be no difference in performance between internal and external GPUs - assuming a proper configuration of course. This is especially true for GPU ray tracing, as the data transfer rate of such rendering is FAR below the PCI Express capacities. Rather than relying on perception, I would benchmark your resuts of the same GPUS internally and externally to see if it shows a difference. If there is significant one, then there is a conflict happening somewhere. So run the tests and please post the results if you're seeing a problem and please make sure to include your hardware configuration, OS, and driver version.

Thanks,

- Phil

[rbcameron1]

Does anyone know if buying a Quadro Plex 1000 IV is worth it to utilize for iRay rendering? Has anyone used a Netstor NA211A - GPU (250watt) to speed up renderings? Its a x8 slot cut open to handle a x16 card, but the host connection is only a x4 pcie slot with 20Gb/s speed. Is that enough to utilize a Tesla C2050/C2075 to its full potential?

[numeless]

Maybe the problem is memory? GTX - 1,5 (3 Gb), Tesla - 6 (8 Gb)
I have a GTX 590 and when I render, both my GPU loaded on average by 30-40 percent.
If the scene is quite simple - loading 90-100%

[numeless]

Try to take the test luxmark 2.0.
The average results for the gtx 580 should be about 415
Here the results for the quadro and tesla:
http://www.luxrender.net/luxmark/res...ser/KnaxKnarke

[SunDevilDave]

It might have come down to the fact that the Tesla card has two DMA engines, allowing for full duplex PCIe communication. This is a feature apparently not available on the GeForce or Quadro products.