Speccing a rendering pc

Jimbo.UK · 24 May 2013 at 15:13

I am currently speccing a machine and looking for some input as to hardware suggestions and priority. I am not asking you to go setup a basket

My purpose of this machine will be to quickly render 3D HD in CAD software. The speed at which it renders is the important factor. It needs to be very fast at rendering as customers will be sat watching and waiting. Making changes and re-rendering.

I am thinking it would be best to make use of CUDA, and prioritize a high spec nvidia card. After that, plenty of Memory and a decent CPU.

Any thoughts?

Thanks.

smilertoo · 24 May 2013 at 16:15

Does your CAD app support cuda rendering? I know a lot of people are impressed with octane render on 3dsmax. For cuda your'e looking at a gtx580 or a new gtx780/Titan, the 680 seemed to be slower for some strange reason. I would actually go for an x79 setup and a 6core overclocked cpu; something i hope to do when ivy-e is released. I also currently have an i920 but at 3.5ghz.

manic_man · 24 May 2013 at 16:24

What are your budgets?

Ideally you could do with a rendering card and a display card.

With a very high budget I would suggest a decent Kepler based card for rendering (GTX 780 or 680, I can guarantee that in most scenarios the 680 will wipe the floor with a 580, regardless of what a few artificial tests throw up, I use both on a daily basis for CUDA calculation of scientific simulations) and a Kepler based Quadro for display (K5000 perhaps?).

If the budget is lower, either a 680 or a 780 on its own with the second being better for CUDA (nearly 500 more cores and a bit more raw performance).

In terms of the base system, you are on the right tracks, really you need as much RAM as you can and don't forget as much deep storage as you can muster. I would also suggest an SSD for OS and application storage, it makes SUCH a difference to the general performance of a system I find. Finally, if you are going to have more than one GPU in there, don;t forget to check that you buy a motherboard which can support 2 cards at PCI-e Gen3 16x (some cheaper models will only have enough PCI-e lanes to allow 2 cards to run at Gen3 8x, which can make quite a difference when dealing with CUDA applications).

And of course, as has been suggested, make sure your rendering packages even support CUDA (or OpenCL I suppose)!

JonJ678 · 24 May 2013 at 16:49

Two options for fast rendering at a "reasonable" price.

CUDA is potentially quick if the software supports it. I'm not sure how well it would scale to multiple cards in one box or whether you can get enough memory on a card to acceptably reduce the overhead of sending data to and from the card.

Second option is having multiple computers wired together, in traditional render farm fashion. You have one normal computer that you and the customer sit in front of. All it does it set up the problem and send it over the network to the render farm then collects the results and draws a picture. That's definitely scalable.