At least it was one reason for NVIDIA to actively support OpenCL, and with all the legal disputes about chipsets and Apple choosing another GPU vendor their interest in developing an alternative to CUDA might have vanished.
Still I would prefer to know about OpenCL’s future on NVIDIA cards. If statements like »NVIDIA’s full weight is behind OpenCL 1.1«… [insideHPC] were up-to-date, we would really have to worry about NVIDIA :-). Momentarily I would not even bet on having OCL 1.0 support with compute capability 3 or 4…
Well it could be an alternative for portable “GPU” computing, unless NVIDIA somehow get upset and kill it(through some sort of patent BS). But why would they, if it is a PTX translator NVIDIA hardware is likely always going to have an advantage since it is designed and revised for NVIDIA hardware.
Anyway, I tend to prefer the OpenCL API as it feels cleaner, ie it has almost everything needed without lots of junk to reduce portability and clarity(and anything lacking can always have an extension). Still some new revisions would be nice to keep it up to date.
I too am waiting for officially released OpenCL 1.1 drivers from nVidia. We are currently using the unreleased cuda 3.1/opencl 1.1 drivers and while we can continue to develop, we can’t sell the product to our clients as they can’t get an OpenCL 1.1 driver (you need to be approved by nvidia to get that old driver) and furthermore, that old driver doesn’t support the new hardware (and well, it has bugs as well).
We feel we are being forced to use AMD hardware… which we obviously will do if need be.
There’s also to say that openCL 1.1, as standard, is still quite behind. many things are not there, many things are still not supported if not with an extension (read writing to a 3d texture, for instance) , and so on.
With this i dont want to say i dont believe it’s not the future.
The future is of course open, and OpenCL even provides the proved working extension mechanism, where vendors can add their own stuff.
I really fear that NVIDIA is again trying to push its new “Cg”, this time making even more people using it before it will fail as de-facto standard, because everybody else run on OpenCL (read AMD/ATI, intel, IBM, imagination technologies)
Not a single word from NVIDIA!?
NVIDIA seemingly does not care about the developers that boasted their hardware
in front of their customers and now look like idiots. We spend a lot of money setting up systems using OpenCL/NVIDIA.
I talked to our customers about how sweet all that is. Now OpenCL 1.1 support starts to become crucial and i feel like
i’m on a sinking ship with all our NVIDIA-stuff. The pre-release driver 258.19 doesn’t compile anymore on our newest kernels,
and should’ve been a temporary solution anyhow.
It basically makes our investments pretty worthless.
I just wrote a support-request asking for information on the matter. I don’t expect to get an answer…
As expected i did not receive any notice from NVIDIA. I canceled our orders and sent out new ones with AMD hardware instead. Maybe NVIDIA will come around again, until then i need to get some work done…
i think the OpenCL v1.1 support is long overdue !!
I think the main reason for this lax attitude from NVIDIA towards OpenCL, stems from the fact that they are already way ahead of the game in terms of OpenCL on GPU and they are faced with any significant contribution. !!
… “Due to the competitive nature of this business, we can not discuss information on unannounced future products or comment on driver feature or date release. We apologize for the inconvenience this may cause.” …
I would’ve considered OpenCL 1.1 announced, but well…
I don’t think so. Check the differences between 1.0 and 1.1 on some dude’s blog. What i am looking for is mainly the ability to query kernel/device for their private-memory-needs and the ability to do region-based r/w/c to image-buffers. The improved multithreading should also come in handy. In my case its more about maintainance that i want ocl 1.1. From 20 Kernels on upwards it becomes hell to count and note the used private bytes in each kernel by hand, and do so again for each change. But i don’t think 1.1 will “simply” speed up your code by 3x.
The next weeks im gonna try out my software on a bunch of amd firepros using their latest ocl 1.1 driver, guess there’ll be a few steppin stones with amd too - but as long as i can develop for 1.1, ill be ok.
It would be bad if NVIDIA decided to drop ocl anyhow, i guess that would be pretty much the death of it. Seeing all the energy going into promoting CloseDA, i fear whatever reasons they once had to promote ocl have vanished.
yes, I see multi-threading as the key differentiator…1.0 does not require mthreading…1.1 requires.
As NVIDIA puts it, all new features appear in CUDA followed by OpenCL.
CUDA 4.0RC made CUDA Runtime multi-thread enabled.
So, OpenCL will follow…
I believe that should happen in 2 to 3 months time… Just my personal view…