NVIDIA Will Soon Probably Introduce OpenCL 1.2 Linux Support

[przemoli]

Yes, absolutely Blender Cycles(Opencl on AMD GPUS)

That is OLD. And CORRECT up until last 13.x Catalyst.


14.x should have most bugs fixed. So You really should check how things stand now, before You bring that argument again.


(For others, AMD had ugly compiler for OpenCL which could and often would produce so big shaders they would not work, if compiled at all. Which was significant impairment for Cycles, which had no such troubles on CUDA)

[Sdar]

That is OLD. And CORRECT up until last 13.x Catalyst.


14.x should have most bugs fixed. So You really should check how things stand now, before You bring that argument again.


(For others, AMD had ugly compiler for OpenCL which could and often would produce so big shaders they would not work, if compiled at all. Which was significant impairment for Cycles, which had no such troubles on CUDA)

Yeah that is OLD and it's not even the first time it was reported that NO COMPLEX KERNEL WORKS ON AMD OPENCL.

But as old as it is, is still true that IT'S NOT WORKING YET.

http://lists.blender.org/pipermail/bf-committers/2014-April/043344.html

Our OpenCL developers are sortof giving up on it... Ton Roosendaal proposes the following strategy now: To offer AMD, Apple, Intel etc support to get Blender and Luxrender installed and tested well. And tell them they can come back to us when it works

Is not working, they stated almost a year ago that it's a priority, is still not working... then they said that they're only going to fix it on GCN cards, and it's not working on GCN cards too... and then they said that they'll fix it FOR GCN FIREPRO CARDS only... but as i don't have a firepro card i'm not sure is this time it's true.

Amd was stalling over a year now... so yeah... is old...

[erendorn]

I lolled on this. Seriously CUDA is a mess right now, which is the reason why almost everything is moving towards OpenCL. No one uses CUDA other than some big companies because they do not work on every kind of hardware instead they work only on Nvidia cards who by the way charges a premium for their cards.

AMD and OpenCL are way better currently.

Tesla cards have better double precision performance and ECC memory, which justify their premium (for those who need that).
Regarding OpenCl vs CUDA, most papers I've seen suggest that hardware tuned OpenCl has the same performance than CUDA.
The difference is that CUDA is already tuned to NVidia hardware, but not usable on other hardware.

I personally went for OpenCl, so that the same code can be used on CPUs, and then GPU can be added when needed.

[johnc]

I lolled on this. Seriously CUDA is a mess right now, which is the reason why almost everything is moving towards OpenCL. No one uses CUDA

Riiiight. Just like how everything is moving towards BTRFS.

[Sdar]

I personally went for OpenCl, so that the same code can be used on CPUs, and then GPU can be added when needed.

Well there's a big problem there... if you are going to do something complex you can choose:

CUDA:
That is going to work only on nvidia cards.

OPENCL
It'll be slower on Nvidia cards.

On amd cards the bigger the kernel the slower the performance ... so it could work so slow that it's faster on a CPU, or if it's just to big it wont work at all. (with small kernels it'll work amazingly fast)

So the question is... Would you start developing something without knowing if it's going to work?

[erendorn]

Well there's a big problem there... if you are going to do something complex you can choose:

CUDA:
That is going to work only on nvidia cards.

OPENCL
It'll be slower on Nvidia cards.

On amd cards the bigger the kernel the slower the performance ... so it could work so slow that it's faster on a CPU, or if it's just to big it wont work at all. (with small kernels it'll work amazingly fast)

So the question is... Would you start developing something without knowing if it's going to work?

Depends.. I'm not very keen either on relying on a solution that I know is not going to work (ie, cuda on anything else). But prototyping will tell

[Sdar]

Depends.. I'm not very keen either on relying on a solution that I know is not going to work (ie, cuda on anything else).

I'll prefer a working opencl too but amd has had enough time already to fix it and the only thing they had achieved is to draw away the developers.

[wargames]

I bought a Nvidia card because life is too short and I can't wait eternally for AMD to provide proper OpenCL support. Yes, CUDA is propietary but it works flawlessly.

[RealNC]

They're preparing for the Maxwell bitcoin miners.

[curaga]

They're preparing for the Maxwell bitcoin miners.

?

GPU SHA256 mining has not been profitable ever since the ASICs came, and for other algos people use CudaMiner. It works exceptionally well, and just recently got contributions from a Nvidia employee.

Nvidia supporting OpenCL 1.2 has nothing to do with that.