AMD to support Nvidia's CUDA technology?

[shiryunaga]

look at this article http://www.techradar.com/news/computing ... gy--612041,

[Leoslocks]

Interesting article and equally interesting response from AMD.

[7im]

IIRC, this idea has been around for about a year now. It may run that way, but unlikely that it will be able to take advantage of hardware specific features, and therefore run slower.

[bruce]

A bit of history:

Intel developed something called Katmai for their Pentimum III. Meanwhile, AMD developed 3DNow! and 3Dnow+ for the same purpose: To enable the CPU to process floating point operations in parallel. Not long after that, Intel developed a final version and called it SSE.

The two developments were initially developed to provide a proprietary advantage to sell their respective hardware. Eventually, AMD decided to pay Intel a licensing fee so that their hardware could be marketed as supporting the more widely adopted SSE.

If AMD eventually does support CUDA (and I know nothing more that what it says in the articles you are reading), I wonder what AMD will have to pay NV to license their proprietary technology. The real question isn't whether Brooke+ or CAL or CUDA or OpenCL is the "best" - - - it's all about competition and ultimately about money.

[Russ_64]

@Bruce, of course you are correct ....... but as folders we want to see CUDA on all GPUs' as it would mean a more streamlined client and cores and I assume better performance on ATi cards.

Currently I get around 3,000 ppd on 4890 with 800 shaders and 5,000 ppd on GTX260 with 216 shaders - so question is would CUDA on 4890 = more ppd than a GTX260 ?

[7im]

No. Cuda on ATI would actually run slower. The ability to run is far from running well, or well optimized for speed.

[mdk777]

Not likely

http://www.theinquirer.net/inquirer/new ... a-run-gpus

AMD's Gary Silcott told the INQ "they [Nvidia] would intentionally damage performance to make Nvidia GPUs run the same app better." Then, perhaps thinking better of accusing Nvidia of hypothetical, yet outright, sabotage, Silcott added "Even if it wasn't intentional, it would not be optimized for our instruction set architecture like our own SDK."

[7im]

AMD's Gary Silcott told the INQ..."...it would not be optimized for our instruction set architecture like our own SDK."

Hmmm... sounds familiar.

[shatteredsilicon]

The problem is that nVidia and ATI architectures are fundamentally different. nVidia is MIMD, ATI is SIMD. In a nutshell:
MIMD:
Pros: Easier to write a decently optimizing compiler. Less dependent on programmer competence.
Cons: Requires more silicon to implement thus more expensive, worse performance per Watt.

SIMD:
Pros: Cheaper, better performance per Watt.
Cons: Writing a decent vectorizing compiler is difficult. So difficult that to the best of my knowledge only two such compilers exist, and only one is for x86 and easily available. Even with a good compiler, reasonable degree of programmer competence (rare and getting more so!) is required to work with the compiler to achieve optimal results.

Now, there is no reason why it shouldn't be possible to write a SIMD capable compiler for CUDA (CUDA essentially being C with type extensions), but this would effectively lead to ATI having to keep up with nVidia's language feature creep, which puts them at a disadvantage when they actually have superior hardware. And as I already pointed out, ATI would have a much harder job in keeping up than nVidia because vectorization is more difficult to do than parallelization. AMD's contribution to vectorizing compilers is based around GCC, and GCC's vectorization capabilities were embarrassingly poor.

This is one of the reasons I have relatively high expectations of Larabee. In terms of raw hardware, it probably won't be as capable as ATI's GPUs, but ICC may well make up the difference.

[Russ_64]

Yeah, I guess we are hoping for a bit too much ........ It still amazes me that after so long, the "superior" graphics hardware is running at a disadvantage in FAH.

[bruce]

Yeah, I guess we are hoping for a bit too much ........ It still amazes me that after so long, the "superior" graphics hardware is running at a disadvantage in FAH.

I don't think that the word "disadvantage" is appropriate. CUDA is superior when you're programming NV hardware but it would not produce superior performance if it were adapted to ATI. The two hardware platforms are really rather different, and the two software implementations are also rather different. I'm not ready to declare that either one has an overall advantage over the other.

Things may change when (and if) OpenCL becomes a reliable alternative, but until then, the jury is still out.

[chumbucket843]

ATi has CAL and brook+. CUDA is for nvidia.

[codysluder]

From what I read, OpenCL might replace both CAL and CUDA.