CUDA 6.5 Is Doing Great For Blender

[magika]

Hah... I would like to make use of my AMD A10-5700 APU's built-in GPU, but too bad. It would be nice if there's an nVidia GPU that does not have video outputs but are very affordable. It will just be a single-slot GPU which I could just turn it off when not in use. I'm not sure if nVidia binary blobs are compatible with Linux kernel 3.16, though.

Blob works up to current 3.17rc1.

Because working on NVIDIA only and being biased is proper now?

I thought stuff that works and yields best results for the task at hand can be considered proper.

[Sdar]

I'm sure it is helpful for their nvidia users, and the person doing this work may not be familiar with, or even care about, opencl.

Blender support Opencl but it's only working on Intel and Nvidia.

AMD Compiler is broken can't compile large kernels, you can compile the Opencl kernel disabling some features but it remains slower than it should be.

[Bucic]

Blender support Opencl but it's only working on Intel and Nvidia.

AMD Compiler is broken can't compile large kernels, you can compile the Opencl kernel disabling some features but it remains slower than it should be.

Any ballpark guess when Blender is going to render using OpenCL on AMD graphics/APUs?

[Sdar]

Any ballpark guess when Blender is going to render using OpenCL on AMD graphics/APUs?

Well, you can disable some features and it'll work on GCN cards... it's slower than an nvidia card... but it works

To use all features we must wait for amd to fix the compiler... don't worry this problem was already reported to amd and they are working on it as a high priority... that's what they said to me 3 years ago...

[liam]

Blender support Opencl but it's only working on Intel and Nvidia.

AMD Compiler is broken can't compile large kernels, you can compile the Opencl kernel disabling some features but it remains slower than it should be.

That's great (not the broken part, obvs).
My comment was focused on vendor lock-in.

http://streamcomputing.eu/blog/2012-09-10/nvidias-industry-leading-support-for-opencl/

Even now, nvidia has only recently moved to opencl 1.2...but at least they support it. So, one api gives you access to the three major vendors.

[kayosiii]

Any ballpark guess when Blender is going to render using OpenCL on AMD graphics/APUs?

Probably not for the current generation of cards (well maybe the latest) I think some of the limitations with AMD's OpenCL implementation are with the cards themselves.

This is true of all other GPU based path tracers I have seen. Hopefully in future both AMD's hardware and OpenCL will be up to the task. Currently the choice between using CUDA and giving some users hardware acceleration and not having working hardware acceleration.

[liam]

Fans or not, when chosing between CUDA that works and does its job (e.g. yielding perfomance wins) or AMD OpenCL implementation that fails to run simple kernels people that care about their jobs would pick CUDA.
Its nice to tell fairy tales about open standards and such and bash proprietary things but when you want to do your job you simply choose proper tools.

Uhm, what was a fairy tale? Cuda only works on nvidia. Opencl works (to varying degrees), on all platforms (don't forget intel, b/c opencl provides a nice speedup even with those gpus)
By your agrieved tone I assume you're the one who works on cuda support in blender. You work for the foundation? If so, and they use nvidia cards, then your work makes a lot of sense.

[Sdar]

A lot of blenders users have AMD cards and Brecht van Lommel (now working on Solid Angle's Arnold render) has spent a lot of time trying to make cycles work on AMD Opencl but at the end Opencl doesn't work on AMD, intel cards are not that fast on renders and Nvidia works better using Cuda instead, so Cuda is at the moment the only hardware acceleration method enabled by default on Blender Cycles.

Anyway I'm not that sure that there's some kind of hardware limitation as kayosiii suggested cause even the AMD CPU Opencl compiler have problems with large kernels, Luxrender that was created in a way to workaround the amd bug can't compile with all features enabled at the same time while the Intel CPU compiler takes only 4.5 seconds so there's just something wrong with AMD compilers.

More info about the CPU compiler problems here:

http://www.luxrender.net/forum/viewtopic.php?f=34&t=11009

[magika]

Uhm, what was a fairy tale? Cuda only works on nvidia. Opencl works (to varying degrees), on all platforms (don't forget intel, b/c opencl provides a nice speedup even with those gpus)
By your agrieved tone I assume you're the one who works on cuda support in blender. You work for the foundation? If so, and they use nvidia cards, then your work makes a lot of sense.

No.
5chars