Nvidia submits its OpenCL 1.0 drivers for certification

[Sharktooth]

Less than a month after offering them to developers registered in the OpenCL Early Access Program, Nvidia has went on to submit its OpenCL drivers for Windows XP and Linux to the Khronos OpenCL Working Group for certification.
Read more... (http://www.tcmagazine.com/comments.php?shownews=26435&catid=3)

... start saying "bye bye" to CUDA... ;)

[deekey777]

Read more... (http://www.tcmagazine.com/comments.php?shownews=26435&catid=3)

... start saying "bye bye" to CUDA... ;)

No, not really. Maybe in 2-3 years.

[Sharktooth]

AMD is also about to submit for certification. OpenCL already lives on Macs, is an open standard and it's multiplatform... CUDA is doomed and Nvidia already knows that.

[deekey777]

AMD is also about to submit for certification. OpenCL already lives on Macs, is an open standard and it's multiplatform... CUDA is doomed and Nvidia already knows that.

C'mon. CUDA is used by thousands developers. They knew, that research and science are the key for GPGPU computing, so Nvidia does a lot for universities (http://www.nvidia.com/page/university_relations.html), there is the Professor Partnership Program (http://www.nvidia.com/page/professor_partnership.html), then their CUDA Center for Excellence (http://www.nvidia.com/page/cuda_excellence.htmll) progam. They support small firms, if they use CUDA. Nvidia's customers are Bull, Cray and NEC, that are using Tesla hardware.
AMD does nothing to make their Stream stuff more public.

Re AMD:
http://forums.amd.com/devforum/messageview.cfm?catid=328&threadid=112397&enterthread=y
We are working with strategic partners currently and providing them with a preview release of OpenCL. A public release of OpenCL will be available in the second half of 2009.

[Sharktooth]

well, that was what i meant by "soon".
however, nothing prevents ppl to continue "using" CUDA. it's just OpenCL is an open standard and will be supported by everyone... not just nvidia...

[MfA]

I think CUDA is going to replace CUDA ... I just don't see the basic architectural assumptions underlying CUDA as it stands now (and OpenCL and DX compute shaders, they are all just minor variations) surviving for long. OpenCL will be the language for the consumerized technology on their old architectures, next gen CUDA will be the language of choice for their newer ones.

Of course once the architectural assumptions underlying CUDA go away all the usefulness of existing research disappear with it :) IMO academic GPGPU is mostly just good training ground for low level programmers, with very little scientific relevance.

[squid_80]

well, that was what i meant by "soon".
however, nothing prevents ppl to continue "using" CUDA. it's just OpenCL is an open standard and will be supported by everyone... not just nvidia...
You might want to look up how much of OpenCL was written by NVIDIA... OpenCL isn't really new, it's basically just CUDA by a different name and everyone has agreed to implement it.

[Sharktooth]

lets see... cuda reminds me of a certain 3dfx's Glide...
the events are evolving in a similar way...

[LoRd_MuldeR]

You might want to look up how much of OpenCL was written by NVIDIA... OpenCL isn't really new, it's basically just CUDA by a different name and everyone has agreed to implement it.

Right. The basic concepts of OpenCL are very similar to CUDA, except that a few terms have been exchanged and that some features, which ATI's hardware cannot support currently, have been made optional. These features are available through vendor-specific extensions, so OpenCL isn't really hardware-independent! Also optimizing the OpenCL code for ATI hardware will be a more tough task for the OpenCL compiler than optimizing it for NVIDIA hardware, because of the great similarity between OpenCL and the underlying CUDA. The good thing is that OpenCL code is compiled on runtime (in order to be executable on different hardware) and hence can be compiled for the CPU as well, if no suitable GPU is available. Of course it will run much slower on the CPU, but at least it does run.

[deekey777]

You might want to look up how much of OpenCL was written by NVIDIA... OpenCL isn't really new, it's basically just CUDA by a different name and everyone has agreed to implement it.

I'm not sure, that this is true. OpenCL, CUDA and ATi Stream are based on scientific projects (BrookGPU eg). So you cannot say "OpenCL is basically CUDA", because it's not.

[LoRd_MuldeR]

OpenCL is not CUDA, but is much more similar to NVIDIA's CUDA than to ATI's FireStream.

For example CUDA's concept of "threads" and "thread-blocks" exists in OpenCL as well. They are only called "work-items" and "work-groups" now ;)

Other features of CUDA, such as "atomic" operations, are available in OpenCL as extensions. So squid_80 was right!

[deekey777]

OpenCL is not CUDA, but is much more similar to NVIDIA's CUDA than to ATI's FireStream.

For example CUDA's concept of "threads" and "thread-blocks" exists in OpenCL as well. They are only called "work-items" and "work-groups" now ;)

Other features of CUDA, such as "atomic" operations, are available in OpenCL as extensions. So squid_80 was right!
From the last Stream SDK:
http://www.abload.de/img/streamt1cz.jpg
http://www.pdc.kth.se/education/historical/2008/Stream2008/index.html/document_view (What's new with R7xx)

http://developer.amd.com/gpu_assets/Shader%20Model%205-0%20and%20Compute%20Shader.pps

[LoRd_MuldeR]

From the last Stream SDK:

Initially and for a very long time ATI's Stream (Brook+) did not implement the concept of "computer kernels" with thread ID's and data sharing at all! Instead they used very basic stream processing (like the original BrookGPU), while CUDA used compute kernels from the very beginning. ATI only hacked all that in lately! So the concept of OpenCL was derived from CUDA and was adopted by ATI much later...

[deekey777]

Initially and for a very long time ATI's Stream (Brook+) did not implement the concept of "computer kernels" with thread ID's and data sharing at all! Instead they used very basic stream processing (like the original BrookGPU), while CUDA used compute kernels from the very beginning. ATI only hacked all that in lately! So the concept of OpenCL was derived from CUDA and was adopted by ATI much later...
Yes, thread IDs and LDS came with RV770 (HW) and in october 2008 with Stream SDK 1.2.1, but only in CAL/IL available.

[LoRd_MuldeR]

Not surprising. These features require changes to the graphics hardware - changes that don't help 3D rendering at all. So you will actually sacrifice some transistors (and hence some 3D performance) for improved GPGPU capabilities. NVIDIA decided to go that way early and it turned out to be a good decision. ATI followed later, probably because the upcoming OpenCL is very similar to NVDIA's CUDA, but was pretty different to ATI's Stream (at that time). Now Stream offers almost the same features as CUDA, but only on the very latest ATI hardware...

[Sharktooth]

that's why opencl is a good thing.

[CruNcher]

In terms of performance/platform independence i would agree it is needed though it still is nothing else then a virtual machine/runtime compiler (similar to java) that false back to the CPU if needed at least that's how i understood Apples initial idea :)
Still as all the other said Nvidia has a big head start over Ati it should be relatively easy to port current existing CUDA code over to platform independent OpenCl and so make now CUDA apps OpenCL aware, and so everyone is gonna benefit from it in the end that is the only thing that count's (lower development costs for different platform distribution) and the performance difference hopefully between a native CUDA application and one that runs via OpenCL won't be that big though i found no information about the Performance differences between OpenCLs runtime compilation Mode compared to a native CUDA implementation especially latency differences ?

[deekey777]

Nvidia'S OpenCL drivers + SDK are out: http://developer.nvidia.com/object/opencl-download.html

(AMD have released their ATi Stream SDK 2.0 beta 3 last week, but without GPU support:)