OTOY Forums

Hi, I write this here, because it is related to speed increase with two GTX 480.

A gain of 1.8 or 1.9 would be great with two cards, but I am a bit worried about something. I explain :

Yesterday, I have received the Cubix GPU-Xpander Pro 2, ordered recently, and today I have connected it. It works fine, and with the GTX 260 that I purchased as an temporary solution, Octane GUI now displays fast and my renders are also faster.

This said, the Cubix box revealed me an unexpected thing: The flashing D4 led reavealed me that my P5K motherboard, bought two years ago and pretendedly equiped with PCIe 2.0 16x by the seller is indead only 1.0. 16x, which is supposed to limit the Cable Link speed transfer to 2.5Gbps instead of 5.0 Gbps. As I had no special need about PCIe before, I never paid attention to that.

Here is my question (I'm also waiting a response from Cubix support, but someone here may already have an answer) :

I don't know what are the real needs in term of speed/bandwidth for graphic cards.
So I would like to know if it will have an impact (slow down) on my rendering time, compared to a PCIe 2.0 16x .

For a moment, I will render (with Octane) on the GTX 260, and later (in the next months when I will have made some more savings and will know if Nvidia will release a "corrected or improved" version of the current GTX 480), I plan to purchase two GTX 480 or similar cards. My fear is that if (for example) One GTX 480 requires the full bandwidth, I could have no speed increase with two GTX 480 despite the double of cores. I have no idea of the used bandwidth for one GTX 480.

Please, could you tell me something about that ?

It is important , because if bandwith is a limitation, it could be more interesting to purchase cards with more memory than cards with more cores.

Obviously, if I get a faster response from Cubix, I will tell you here.

I belive radiance somewhere wrote that they did test with cubix expansion box with two gtx480 in it, connected to one mobo 16x PCIe slot with no speed impact over two of those cards running in two separate x16 slots. Only "drawback" was slight slower loading time for cubix(few sec)

Cheers,
n1k

@n1k :Thank you !

I have just received a similar response from Cubix.

It is not true for all Cuda applications, but Octane traffic on the cable link between the expansion box and the computer seems to be effective only during the loading of the scene.
however, as the rendering is displayed in real time on the monitors, I think that there is still a transfer on the link, but probably with low bandwidth consumption.

For animations though, I think that having a 1.0 PCIe slot instead of a 2.0 may increase the loading time of the frames...