Why do we not use openCL on Nvidia cards? Most of the newer Nvidia cards have openCL support. I’ve seen people talking about how much better OpenCL is to use.
Zach
1 Like
Well, if you set a certain environment variable, OpenCL does actually work on Nvidia.
But CUDA does the same thing faster. So it’s pointless to use OpenCL on Nvidia beyond just experiments and testing.
1 Like
I guess I misunderstood what I read (or the people that wrote it were mistaken). They were saying that OpenCL was better and faster than CUDA. I also read that OpenCL is able to use the GPU and CPU at the same time.
1 Like
Well yes, it can. However, the performance gain is not really worth the feature loss - Cycles’ OpenCL kernel is a bit of a stub, missing most of the more advanced features because there just hasn’t been a lead platform where OpenCL is preferable to all other solutions - the CPUs have their own dedicated C++ implementation, and Nvidia GPUs have CUDA.
AMD GPUs which should be the lead platform for OpenCL just don’t work at the moment because the AMD OpenCL compiler has been optimized for much smaller kernels to the point where it just can’t cope with something as big as Cycles (it tries to inline just about everything to get rid of function calls, but since Cycles is so big, it runs out of RAM before it can complete the task). If you pick up one of the older builds (around 2.61 or 2.62 i think) and have a lot of of RAM (not sure, either 24 or 32GB )you just might get it to render something after waiting a long time for the compiler to finish.
But then since you’re asking about OpenCL on Nvidia, this probably isn’t of much interest to you.
Oh I see. Thanks for the reply. I was wondering about this for a while!
Zach
Cycles’ OpenCL kernel is a bit of a stub
As far as I know, the OpenCL kernel is now feature-completed (compared to CUDA). It’s right now interesting mainly because it renders significantly faster than a (non-mingw) build on Windows using CPU. It also supposedly(?) lets you combine CPU+GPU, however that combination does not show up for me.
Oh. When did that happen? Anyway, thanks for correcting me.
DingTo confirmed so. However, seems like you need to change some hardcoded settings to get “advanced shading” to work, still.
Hi, you can enable this since rev57054 with:
export CYCLES_OPENCL_TEST=all
in a CMD before you start Blender from there.
It renders much slower as CUDA and may help only if you have a strong CPU but a small GPU.
On my system (specs in signature) CUDA is 3 times faster as OpenCL with CPU and 2 GPU.
http://projects.blender.org/scm/viewvc.php?view=rev&root=bf-blender&revision=57054
Cheers, mib.
I downloaded and tried r57264 from graphicall.org to see if it works. It doesn’t show openCL in the user preferences where you pick CUDA. Its supposed to have the patch applied to it so that you can use openCL.
Zach
mib, could you explain this further?
I found this on google. The “export” command is linux only.
Zach
any luck for windows users?
Bao2 says you can do this:
but I don’t understand what he means by “make a link”
Zach
I figured it out
in the command prompt, you need to navigate to the blender directory, then use the set command in place of the export command.
managed to get it mostly working, still doesn’t seem quite the same functionally as cuda. i think there might be some issue in the mix shader. still, an exciting development!
Hi guys, in windows it call set not export, sorry.
To activate Intel CPU you need also Intel SDK:
http://software.intel.com/en-us/vcsource/tools/opencl-sdk
Cheers, mib.
In case you aren’t comfortable with command prompts, you can just set the environment variable in your system settings:
when they talk about cpu + gpu im pretty sure its for intel cpus with the integrated gpus… for now
i may be wrong
:3
LordOdin, no, it is for Intel CPU and Nvidia GPU.
I cant get integrated gpus to run on win 8 but this is may because I use Nvidia for display and the the intel video driver is not loaded.
Cheers, mib.
looking at it further, it appears that MIS isn’t working under openCL