OpenCL 1.1 driver, 8 months and waiting...


Does anyone have any info on when OpenCL 1.1 will be supported by NVIDIA? (And I mean OFFICIALLY supported in a public driver - not some pre release beta buggy crap). I’ve been working with an algorithm that uses OpenCL 1.1 features, and the code works fine on a AMD Cypress (5870) card. But using the beta pre release buggy crap driver from June, I get all kinds of weird problems. The card I am testing with from NVIDIA is the Tesla 2050 (Fermi), and it is shared with other developers that are testing CUDA apps and of course they prefer to have a modern driver for their CUDA codes. Should I just write off Nvidia’s OpenCL support, or is this something they are actually addressing (beyond lip service).



I’m also interested in this. The RC driver that was released on June 2010 met conformence on July 2010, acording to the OpenCL section in the CUDA zone.
What keeps the driver and respective toolkit from being released publically? Would be nice to get an NVIDIA official response on this matter.

Best regards,

Liad Weinberger.

I am also working on a OpenCL project and i would also like to have an idea why the driver and SDK is not out yet?

Strange enough, there is not a single word from NVIDIA about its OpenCL 1.1 implementation, and I am googling for it regularly.

For quite some time, NVIDIA made a lot of commitments to OpenCL. Now, NVIDIA has a compliant OpenCL 1.1 implementation for more than half a year, and we hear NOTHING. Personally I could make up only a single interpretation of this fact that makes sense, and its not a very good one for the OpenCL community…

I get the feeling that nVidia is not very committed to OpenCL. After all, it is a competitor to CUDA. Just look over the OpenCL forum and anyone can see their CL implementation behaves in obviously broken ways. Not being committed to open standards is dangerous. The only way to compel NV to support CL better is to hope competitors like AMD and Intel make better CL implementations and take NV’s business.

I have expressed my disappointment several times in this forum and there is no response from NVIDIA. Now the CUDA 4.0 RC has been announced and will be ready for registered developer later this week. I ask myself where is the future of OpenCL and should I discard it?

Same disapointment for me …

But my new hope - strangely - come from the new Apple MacBook line … without any nVIDIA cards…means that the software developers who wants to have GPU acceleration on Mac will try OpenCL so … more pressure on nVIDIA, no ?

The OpenCL community needs big software to use OpenCL !

Well I am glad others are frustrated too. My main annoyance is the fact that Nvidia made a big splash about being OpenCL 1.1 compliant, when in fact they are not. They are only OpenCL 1.1 compliant in a closed beta environment, which is worthless if you are developing apps to be deployed in the here and now. I also do not understand the delay, and the conspiracy theorist in me is left to assume this is a strategic decision to bolster CUDA. The downside to this (at least from my own apps), is that AMD is running rings around Nvidia in the OpenCL market.

I see a much more severe problem - M$’ DirectCompute is just waiting to fill the widening gap. Our industry partners typically do not care of a 10 percent or even a 50 percent difference in performance (NVIDIAs C4CUDA and OpenCL implementations are similar in many, many cases, anyway), but they are really frightened to be tied to a single supplier (or in most cases: to another supplier besides M$ with its OS).

But we might also be lucky and CUDA4RC provides an OpenCL 1.1 implementation and proves us wrong, we’ll know by the weekend…

There has been no announcement of this kind. The CUDA4RC will also be closed only for registered developers till the 4.0 release is out, so in the meantime this means that even if it does support OpenCL 1.1 on top of it, it wont be publicly available.

I must admit, this boggles me, and to be honest, even upsets me. I’m a CUDA enthusiast as much as I am an OpenCL enthusiast. For projects where I know for sure that the underlying hardware will be NVIDIA’s GPUs that will do the number crunching, I’d usually prefer CUDA, as it is more tuned for the hardware. However, this isn’t always the case, and some projects need to support multiple vendors. OpenCL 1.1 offers some great ways to improve the robustness of the code. Having no OpenCL 1.1 support for NVIDIA chips, makes the task of maintaining specialized versions of kernels and host invocation wrappers even more tedious.

I completely agree that NVIDIA is taking way too much time on releasing official OpenCL 1.1 drivers, or at least new beta drivers which would support their current (or older) lineup. As an example we just bought a gtx460 for a new computer. Obviously as they are kinda old fermi cards we kind of assumed automatically that the beta drivers would work with it. But of course they do not. It’s getting tedious to search for a NVIDIA GPU with OpenCL 1.1 support.

On the other hand ATI drivers tend to crash all the time so our only choice is to carefully pick the NVIDIA cards that do have beta driver support if we want to use OpenCL 1.1. So please, release us from this torment :)

So CUDA 4.0 is around the corner, but still no word on shipping OpenCL 1.1 support. :(

Little surprise. I’ve downloaded the CUDA4.0RC DevDriver and Toolkit. Nope, no OpenCL 1.1 support.

Anyone an idea about OpenCL 1.1 support in MacOS X 10.7 Lion? (If you know and happen to be an ADC member, make sure not to break any agreement by disclosing that information…)

I finally found the answer myself in the insanelymac forum. So Apple’s Lion Preview has an OpenCL 1.1 runtime, but the contained NVIDIA drivers provide 1.0 support only (for now).

NVIDIA likes to claim that nobody is using OpenCL to keep momentum for CUDA going, but this couldn’t be further from the truth. They aren’t likely to budge from this viewpoint for the time being unfortunately.

Sigh, which is just going to push people into using DirectCompute or stick with OpenCL 1.0 if they want GPU portability(this is probably not the kind of thing which would convince many to switch to CUDA if that is what they want).


When have they claimed this?

Maybe claimed was too strong of a word. But their actions and lack of support for OpenCL show that they think they can just get everyone to use CUDA and be locked into their hardware.

I’m wondering if the primary driving force behind OpenCL is/was selling hardware to Apple. Both for GPU’s and CPU’s.

gpuocelot now supports AMD GPUs so perhaps this will be come a viable alternative to OpenCL.

What is needed is a good open sourced OpenCL 1.1 compiler.