GPU Rendering Box?

[johnharper1]

The advice from the 3D guys is it's better to build a box with, say 4 GPUs, instead of two boxes with 2 GPUs each. They say it's more cost effective, but I'm having trouble finding a MB with 4 x16 slots? When the pciex speed is reduced with multiple GPUs, does that effect efficeincy?

 

Lastly, I'm trying to wrap my head around the PSU requirement and other factors for this type of setup. Can anyone offer some advice? TIA

[Nikolaos M]

You won't find a single socket motherboard with true x16 lanes for more than two pcie slots, and those who claim to have x16 for more slots use PLX switches, which artificially multiply the lanes, assigning bandwidth to each gpu depending on the actual load at each point in time. But even x8 per gpu would be just fine in terms of actual bandwidth. It would be more than enough. So, don't worry about that.

 

Imo, it comes down to how you're planning to deal with the cooling of the cards, when most of comercial gpus are dual slot in width. There is a single slot 1070 coming out later this year, this one in particular http://www.anandtech.com/show/10926/galax-shows-off-single-slot-geforce-gtx-1070-graphics-card. It would be a nice solution for a multi-gpu setup.

 

As for the psu, get one with the highest wattage, efficiency and quality you can afford, and make sure it has a decent headroom in Watts, because gpu rendering is not only about average consumption, but you have to take into account that during rendering there are spikes occurring and this situation tests the psu's stamina and quality. Solid choices imo are these ones:

https://pcpartpicker.com/product/v7PzK8/evga-power-supply-220t21600x1

https://pcpartpicker.com/product/LbtWGX/evga-power-supply-220p21600x1

They have a 10 year warranty too.

 

Another factor of course is the chassis (very important when it comes to gpu temps), and the gpus Vram. Most gpu renderers are limited by Vram (I think some of them like redshift can use the system's RAM too, and Cycles is scheduled to adopt this feature soon).

Edited January 24, 2017 by nikolaosm

[johnharper1]

Thank you.

[Dimitris Tolios]

Even 1-2x "True" PCIe 3.0 bandwidth available per card is enough to access 100% of the cards. The 16x "thingy" is a myth. Was for 2.0, is for 3.0. Bottlenecks exist, PCIe is just too far down the list.

 

By the time cards will be close to pushing the limits of the 3.0 standard in RT multi-GPU configurations (so far Gaming was more taxing than GPGPU, at least in the PCIe bandwidth demand), you will start seeing PCIe 4.0 platforms which are already documented and ready.

But people in forums and FB groups and whatnot love to act as if them as players are bigger than the game, and have figured out stuff better than the engineers setting the operational envelopes and alternative upgrade paths a decade (or so) in advance.