[linkedin] Better OpenCL support in NVIDIA's CUDA SDK petition

[Final8ty]

Introduction: CUDA, OpenCL, GPGPU

CUDA and OpenCL are techniques to program GPUs, where CUDA only works on NVIDIA GPUs. CUDA and OpenCL are often referred to as GPGPU- General-purpose computing on graphics processing units. Various researchers have concluded that with OpenCL and CUDA comparable results can be reached.
Besides GPUs, OpenCL is designed to also work on DSPs, mobile processors, FPGAs and modern CPUs. It is not performance-portable, but all techniques learnt can be used on the various platforms.
OpenCL is very promising, but like with all open standards it needs support of the developers and goodwill from the participating companies.


The problem

Nvidia is not including OpenCL samples in the latest CUDA SDK and has removed profiler-support for OpenCL, and instead focusing more on their proprietary CUDA. The reason is simple: every developer who chooses CUDA over OpenCL, is limited to NVIDIA hardware.
As a Khronos member with an excellent record in implementing and promoting standards like OpenGL, this is a surprising and even unacceptable behavior from Nvidia.
OpenCL developers need a full-blown SDK (such as OpenCL samples and aprofiler), so the potential and limitations of NVIDIA GPUs can be learned. Also industry standards like OpenCL help in building up a bigger market for GPU computing, and will be beneficial to Nvidia in the long term.
What I aim by signing this petition
By signing this petition, I request Nvidia to put back the OpenCL samples and profiler in their latest CUDA SDK. By this, the choice is put back to the developer, as it should.

http://www.linkedin.com/groups/Petition-put-back-OpenCL-samples-1729897.S.149050851

If you would like to sign please visit - http://www.ipetitions.com/petition/o...in-cuda-5-sdk/

Someone with a twit account should tell Roy@AMD to pull there own finger out on OpenCL.

https://twitter.com/amd_roy

 

[Vesku]

Might want to indicate it's a "petition".

 

[sontin]

Really? Forcing nVidia?

Buy AMD or Intel or some other GPU vendor stuff.

 

[Imouto]

Force AMD to make a frigging OpenCL compilant driver. Ppl want consumer level applications with OpenCL enabled and devs are forced to go CUDA because AMD is unable to make its drivers work properly. You may be mad at Nvidia about its antics but AMD is really pissing off developers.

Summary:

- Nvidia is crippling its OpenCL performance.
- AMD is helpless about OpenCL, if you want any complexity forget about it.
- Intel is expensive as f*ck.

 

[Jodell88]

Force AMD to make a frigging OpenCL compilant driver. Ppl want consumer level applications with OpenCL enabled and devs are forced to go CUDA because AMD is unable to make its drivers work properly. You may be mad at Nvidia about its antics but AMD is really pissing off developers.

Summary:

- Nvidia is crippling its OpenCL performance.
- AMD is helpless about OpenCL, if you want any complexity forget about it.
- Intel is expensive as f*ck.

What level of complexity are you looking at? GPU rendering can be done in OpenCL and I would consider that as complex.

 

[ultimatebob]

Why do I feel like most of the people signing this petition are Bitcoin miners?

 

[Imouto]

What level of complexity are you looking at? GPU rendering can be done in OpenCL and I would consider that as complex.

Well, now that you mention it please give a solution to Cycles rendering engine developers for their OpenCL implementation in AMD cards. They were forced to go CUDA because of issues with the AMD OpenCL compiler.

Even the Luxrender team is having problems with AMD drivers when their renderer is reaching maturity. Luxrender is working beautiful on AMD gear but they're having trouble lately.

When you start adding features to your render engine its complexity skyrockets. Volumetric lights, caustics, new shaders, procedural textures... With AMD compiler developers run into problems way before than CUDA.

 

[Jodell88]

Well, now that you mention it please give a solution to Cycles rendering engine developers for their OpenCL implementation in AMD cards. They were forced to go CUDA because of issues with the AMD OpenCL compiler.

Even the Luxrender team is having problems with AMD drivers when their renderer is reaching maturity. Luxrender is working beautiful on AMD gear but they're having trouble lately.

When you start adding features to your render engine its complexity skyrockets. Volumetric lights, caustics, new shaders, procedural textures... With AMD compiler developers run into problems way before than CUDA.

Yes, I know the problem with the AMD compiler consuming ungodly amounts of RAM when compiling the kernels and then crashing. Brecht and the Blender dev team encountered it early on in Cycles development and couldn't find a work-around. That is why the CUDA implementation is so far ahead in terms of features. However, if AMD could fix the compiler I'm sure the Blender devs would prefer to use OpenCL.

The Luxrender devs only encountered the bug in deveploing SLG4 and its integration into Luxrender itself. At the end of it all it's not the language but the tools.

 

[Vesku]

Well, now that you mention it please give a solution to Cycles rendering engine developers for their OpenCL implementation in AMD cards. They were forced to go CUDA because of issues with the AMD OpenCL compiler.

Even the Luxrender team is having problems with AMD drivers when their renderer is reaching maturity. Luxrender is working beautiful on AMD gear but they're having trouble lately.

When you start adding features to your render engine its complexity skyrockets. Volumetric lights, caustics, new shaders, procedural textures... With AMD compiler developers run into problems way before than CUDA.

A good chunk of the issues are that CUDA has been established for a longer period of time. Which is really where these complaints are revolving around. They want the maturity of CUDA but the shared standards multiple hardware targeting of OpenCL. That's going to take time and I bet some of the stuff they are bummed about regarding current Nvidia OpenCL support is not deliberate dragging of feet but merely the reality of man hours.

Although this in particular "Nvidia is not including OpenCL samples in the latest CUDA SDK and has removed profiler-support for OpenCL" is perhaps Nvidia not wanting to spend the man hours ($) to support OpenCL.

 

[Jodell88]

A good chunk of the issues are that CUDA has been established for a longer period of time. Which is really where these complaints are revolving around. They want the maturity of CUDA but the shared standards multiple hardware targeting of OpenCL. That's going to take time and I bet some of the stuff they are bummed about regarding current Nvidia OpenCL support is not deliberate dragging of feet but merely the reality of man hours.

Although this in particular "Nvidia is not including OpenCL samples in the latest CUDA SDK and has removed profiler-support for OpenCL" is perhaps Nvidia not wanting to spend the man hours ($) to support OpenCL.

I remember OpenCL performance dropping around 50% in the late 200 series drivers and it has not picked up since.

 

[Vesku]

I remember OpenCL performance dropping around 50% in the late 200 series drivers and it has not picked up since.

My guess is Xeon Phi has Nvidia concerned so they are trying to sandbag their CUDA customers. Timing would make sense.

http://software.intel.com/en-us/blo...cing-opencl-12-for-intel-xeon-phi-coprocessor

 

[Imouto]

I don't see myself going back to AMD for the time being. I tried hard to get accustomed to Luxrender but its lack of a node editor (will be fixed soon) and beta-like status of GPU rendering make it a hassle. Cycles integrated into Blender and a $200 used 3gb 580 is all you need for cruise control to awesomeness.

I'm sure I'm not the only one that ditched AMD for this reason and they should focus on open source software like Blender to show off features and performance/price ratio to commercial soft devs. The computing power is there, AMD just need to get their OpenCL stuff together.

 

[VirtualLarry]

Although this in particular "Nvidia is not including OpenCL samples in the latest CUDA SDK and has removed profiler-support for OpenCL" is perhaps Nvidia not wanting to spend the man hours ($) to support OpenCL.

The more cynical among us might think that NV is actually trying to kill off OpenCL, before it becomes more popular than CUDA.


Though, NV isn't 100% of AMD's problem. Part of AMD's problem is AMD's own support for OpenCL.

 

[Final8ty]

Someone with a twit account should tell Roy@AMD to pull there own finger out on OpenCL.

https://twitter.com/amd_roy

 

[sontin]

The more cynical among us might think that NV is actually trying to kill off OpenCL, before it becomes more popular than CUDA.

And AMD is doing the same with HSA, too.
Funny that they get no flak about it. :sneaky:

 

[zlatan]

And AMD is doing the same with HSA, too.
Funny that they get no flak about it. :sneaky:

HSA is an open foundation. If NVIDIA want to support it, then they are welcome.
Also HSA is not an OpenCL or C++AMP alternative, it just expand these platforms.

 

[sontin]

Sure, it is. And AMD is not producing real products. D:

AMD will only push standards which will run at best on their hardware. :thumbsdown:

 

[zlatan]

Sure, it is. And AMD is not producing real products. D:

AMD will only push standards which will run at best on their hardware. :thumbsdown:

Ofcourse it run best on the GCN architecture. But they don't force the competition to not create a better hardware.