Tweet
I'm trying to understand the benefit of additional video cards on GPU rendering, if I have one card that gives me a 30 times increase over regular Vray-RT what will 2 cards, 3 cards, 4 cards and so on give me? Is the speed increase truly incremental or is there some overhead associated with coordinating the GPU's and at what point does the benefit of adding cards stop.
-
-
I have not tested it myself, but most of the folks I have spoken to seem to think that the speed will scale in a fairly linear fashion, at least in a single-machine configuration. This is probably due to the parallel nature of the cores on the GPU, and the way they access the memory, which will have the entire scene loaded. A concern here is that you must use some GPU cycle for the display, and that can slow things down a bit as you balance the GPU load between display and openCL rendering.
It looks like the biggest concerns will be in multiple machine setups where the passing of data between machines on the network may slow things down a bit. Someone here has 9 480s (I believe in seperate machines) so perhaps they might let us know how the speed is scaling.
-Alan -
That's exactley what I want to know, the setup I'm looking at building will be 5 dual processor machines each with at least one 480 in it. If GPU rendering is anything like distributed rendering with each additional machine you add you see smaller and smaller benifits; if I'm using 5 480's and adding a 6th card only gives me 5% more power then it might not be worth it.Originally posted by Alan Iglesias View PostComment
-
Drop the 5x dual processor machines and get this http://www.evga.com/articles/00537/
Stick 3x NVIDIa, or w8 for ati support and 4x ati dual gpu. OC ur xeons to 4ghz and enjoy ur time with vray. 1/2 machines like that will smoke.Comment
-
Yes, I understand the concern regarding diminishing returns. Vlado can of course explain this much better than I, but I believe that because RT is sending a Vray-specific file to each GPU which now participates in the progressive rendering using a stand-alone version of the Vray renderer, there is much less traffic between machines as compared to classic DR spawning in which a Max file is created and then distributed to nodes running Max and then contributing buckets to the render. So I'm hoping that with RT you can use significantly more nodes with multiple GPUs before seeing an appreciative slowdown.
It would be great to get some scaling numbers from the fellow with the nine 480s. Here's the thread...http://www.chaosgroup.com/forums/vbu...mage-sequences
-AlanComment
-
I'm also a little confused about which card to buy, the Tesla C2050 costs around $2,500 it has more memory than the 480's but about the same number of cores and I could get 5 480's for that price. I would think 5 480's would be much faster than one C2050 with memory being it's only advantage, so why is it so expensive?Comment
-
Bcause it is made for scenes that have 50000 + objects lots of wireframe and so on. Its not about the speed. Its about having LOADS of geometry and having 100 fps in viewport.Comment
-
So from what I've gathered a 30 times increase over normal RT (CPU) rendering is based on the test done by chaos using 3 480's, if that's correct then with the addition of each card to the system you get a ten fold increase in speed. If you have 10 cards that should translate into 100 times faster than what RT (CPU) alone can do, Vlado is that right?Comment
-
I doubt that you can add 10 cards in one machine, but in any case the increase will likely not be perfectly linear.
Best regards,
VladoI only act like I know everything, Rogers.Comment
-
Later today I am going to upgrade my quadro 5600 with a quadro 5000 SLI'd with a Tesla C2050 I am preping benchmarks with the 5600 now, will post the results soon.
-DaveComment
-
@Dadal. Please do not show me such things. Makes me very upset... I just looked at my bank account.
@Vlado. I really think we need a hardware section.Comment
Comment