Is Radeon pro duo 32GB worth it for cycles?


(MALLEKMIROX) #1

Hi guys,

I am considering a workstation slowly for blender cycles rendering that has 2 X Radeon pro duo 32gb graphics cards, do you think that OpenCL is future proof for blender ? or CUDA will still be better ? , I have a workstation with 2 X 1080 ti for the moment and I want to cut my render time more, utilizing another computer, do you think that 2 Radeon duos > 2 1080 TIs ?

I am kinda hesitating if I should try this built since I never tried AMD cards, if anyone has one of these or know any information that would help a lot.

Thanks!


(bernardothecat) #2

I would wait for reviews/benchmarks of the new Radeon VII.


(Grzesiek) #3

Radeon Pro Duo 32GB is basically 2x RX 480s with extra ram.

1x 1080ti, in rendering is almost 2x performance over RX 480.

I would definitley wait and see what that Radeon VII has to offer, and with only 1 month to go, do wait.

As for OpenCL, that will continue to be supproted for long as Blender is open source and technically it does support both Nvidia and AMD (and soon Intel) GPU’s. I do not expect this to be wiped anytime soon.


(BigBlend) #4

Even if OpenCL is supported it is not great for quick rendering as it needs to compile for 30s-60s just to start the render. Also the hybrid solution with cpu+gpu, the opencl solution is way behind CUDA solution.


(Grzesiek) #5

True. The complie part… that is the worst. But that is with AMD’s OpenCL compiler issue. It wasn’t able to deal with the single monolithic kernel for Blender, Cuda was. AMD chipped in developers to split the kernel to chunks and now it “works” but as you said… not great.

Still looking at ProRender from AMD which is OpenCL driven, this does not happen as such it seems to indicate that simply work on blender mega kernel is required to fix this issue.

Still 30-60 seconds. i guess with large complex scenes? Guessing it is very CPU dependent. I usually get 10-15 seconds, but mainly models with textures including glass. so not overly complex.


(Grzesiek) #6

Ahh they hybrid solution. can you give me more input how well CUDA works with cpu+GPU?

In OpenCL, for me, it works good, I do get noticeable improvements. But will definitely need to test it in B2.8b


(BigBlend) #7

In CUDA you can use 16*16 tiles without slowdowns. OpenCL can’t. Also there are builds for CUDA that are like 40% faster than the official Blender.

The link E-Cycles - Faster cuda rendering


(Grzesiek) #8

yup, seen the e-cycles one. we’ll see how that goes, well probably more so… any speed improvements CUDA or OpenCL are a good thing.

as for 16x16… will get back to you but in 2.8 and 2.79 latest builds, that didn’t seem to be an issue. Still i’ll confirm and post my testing.


(kabu) #9

About opencl vs cuda…
https://news.ycombinator.com/item?id=16892595


(Grzesiek) #10

Id be eager to get input more from the Blender Developers. they are the only ones who can provide clear information on this, else it is just guess work


(mat) #11

OpenCL uses the split kernel. The split kernel needs more rays at once to feed the GPU, which means you need bigger tile size. Their are 2 ways to get OpenCL to work well with small tiles:

  • render many tiles at once, is in plan, Brecht said it will come, when is unclear
  • get the mega kernel to work properly on AMD hardware, will certainly never happen…
    So GPU+CPU on AMD hardware may be an option in the future, but not yet.

Regarding the kernel build time, I could work on that if enough people show interest to get it in E-Cycles. As usual, it would then be submitted for inclusion in master after a year.