Octane performace with PCI-E x1

[Jazbee]

Hi

Just want to ask if anybody knows how plugging GPUs to different PCI-E slots affects Octane performance. With regular GPU performance (games, etc.), it's a no-brainer, it slows down significantly. But what happens when only the number of CUDA cores is taken into account in terms of performance (like with Octane)?

I actually know a guy who has 6 GPUs attached to one motherboard using the PCI-E extender cables (and a VERY custom case). Only one GPU is attached to PCI-e x16 and others are plugged into the PCI-E x1 slots (it's true, I saw the photos). He claims he doesn't see a performance penalty in Octane, but I'd like to know if it can be confirmed on a technical level.

Thanks

[kavorka]

I have read that the only difference is the speed of loading everything to the GPU memory. So the only difference you would see is load times, but rendering speed shouldn't see any change.

If you are not loading large scenes, you probably wouldn't even notice it.

[Jazbee]

I have read that the only difference is the speed of loading everything to the GPU memory.

Thanks! Do you remember where you read about it? Was it someone from OTOY or just general discussion?

[kavorka]

Ya, pretty sure someone from Otoy mentioned it.

[Obizzz]

Is this because the entire scene is uploaded to the video card memory?

[kavorka]

Ya, my understanding is that it loads everything onto the GPU memory using the bus speed, then once its there the bus isnt used much except for maybe saving the image out, ect.

I'm sure of the developers could break the whole thing down. If you do large scenes, some of mine on x16 will take a good 2 min to load all of the geometry and textures. Assuming the x16 is actually 16 times faster than the x1, that means that scene would take 32 min to load. I've never used lower than x8.

[BorisGoreta]

This is not correct, when your scene is uploading all kinds of things are happening, not only transfer to GPU. PCIx1 is 500MB/s and PCIx16 is 8GB/s or even more.

I have 5 GPUs in a network node which means they are connected via network cable which is 100MB/s, five times slower then PCIx1 and I have to wait for only a couple od seconds to 10 seconds more for it to start rendering compared to other GPUs which are directly connected using PCIx16, if you are rendering animations then for every additional frame you would have a delay of a fraction of a second because only thing transferred to PCIx1 GPUs is the position of a camera.

All in all PCIx1 is very good. Don't worry.