AMD Gets CUDA Compiler, Going All-In on HPC in 2016-2017

[dalekphalm]

This does not enable game works, it gives AMD the ability to use HIP, Gameworks is PTX only. Nvidia would have to write/translate Gameworks to run on AMD cards on their own accord.

Thank you for clarifying for others. But I did not claim it would work with GamesWorks.

Frankly I don't believe they got a license at all yet. The evidence doesn't point to that.

I would love to be proved wrong.

If they got a real actual CUDA license (which I don't believe), they could bake in GamesWorks support in future GPU's.

Or possibly have some sort of hardware translation to enable PTX/CUDA C to run. But maybe that's impossible with current GPU's?

[Dabombinable]

I just noticed a lot of people don't seem to understand what AMD actually got a license of.

Cuda is generally implemented through PTX, Nvidias proprietary intermediate code.

AMD cards in no way shape of form can support PTX, meaning programs compiled with NVCC won't work. (They will have to be translated or rewritten from the source)

So legacy support is impossible unless the applications are recompiled and future application support is sketchy. 

AMD will/now have the ability to compile/read in a subset of CUDA called HIP (Still CUDA).

A source-to-source compiler will be available but itll be up to devs to use and debug the exported code.

 

ALSO NVIDIA GAMEWORKS and PHYSX ARE PTX!

 

 

THIS IS NOT FOR GAMING!

And if they redesign an architecture to support PTX?

[Tomsen]

No, AMD wont directly support CUDA. Would be rather ineffective.

This is to ease out porting meanwhile still bringing higher optimization than directly supporting CUDA.

 

http://www.phoronix.com/forums/forum/phoronix/latest-phoronix-articles/834978-amd-working-on-cuda-source-translation-support-to-execute-on-firepro-gpus/page2

[Notional]

@patrickjp93 Could you please change that decietful headline already? AMD did not get a CUDA license. Nothing in any sourcing surrounding this news confirms, or even hints at such an occurrence. In fact, the big sources even discusses how this could violate CUDA IP (although that seems like BS.

 

Fact of the matter (and yes this is a fact), is that AMD hardware cannot run CUDA code, and all this is, is a new compiler, that helps developers translate CUDA code to C or similar languages, that are not proprietary in the same way.

[Guest]

And if they redesign an architecture to support PTX?

Nvidia didn't give them the spec for PTX, or a Cuda License. Plus why would Nvidia? 

[don_svetlio]

Nvidia didn't give them the spec for PTX, or a Cuda License. Plus why would Nvidia? 

Because Nvidia are generous and noble and help others? /s

[Nup]

 Noob  question here; can a device use cuda regardless of it's structure? Is it mostly to do with kernals or does it have a level of hardware implementation? Or is it simply a licence to buy? 

[WarlockTheWise]

 Noob  question here; can a device use cuda regardless of it's structure? Is it mostly to do with kernals or does it have a level of hardware implementation? Or is it simply a licence to buy? 

There are loads of things which are only supported on newer hardware, like dynamic parallelism for example.

[Prysin]

And if they redesign an architecture to support PTX?

then they would definetively get sued....

 

CUDA and everything tied to it is Intellectual Property and licensed and prolly also patented, by Nvidia... the HIP system alone is borderline enough to sue over.

[Gibs960]

Am I right in thinking that this will close the gap between AMD and nVidia cards?

[dalekphalm]

Am I right in thinking that this will close the gap between AMD and nVidia cards?

Not really, and definitely not in video games.

 

At best, it will make it easier for GPGPU Compute developers to port their CUDA programs to the AMD platform (Using HIP?). The AMD platform will still not be running CUDA. The compiler is just taking the CUDA code and translating it into code that AMD supports, then compiling it into an AMD compatible program.

[Gibs960]

Not really, and definitely not in video games.

 

At best, it will make it easier for GPGPU Compute developers to port their CUDA programs to the AMD platform (Using HIP?). The AMD platform will still not be running CUDA. The compiler is just taking the CUDA code and translating it into code that AMD supports, then compiling it into an AMD compatible program.

 

 

This is probably a stupid question but, what's the point in this then if it won't affect performance in any meaningful way?

[Notional]

then they would definetively get sued....

 

CUDA and everything tied to it is Intellectual Property and licensed and prolly also patented, by Nvidia... the HIP system alone is borderline enough to sue over.

 

What are they supposed to get sued for exactly?

 

This is probably a stupid question but, what's the point in this then if it won't affect performance in any meaningful way?

 

This entire topic has nothing to do with gaming or consumer in any way. This is server/workstation stuff. The entire point of AMD's focus here, is to help programmers of CUDA based software, port their software/programming to a different non proprietary language and then compile it to something that will run really great on AMD hardware. The point is to gain market share in the CUDA based compute markets.

[Gibs960]

What are they supposed to get sued for exactly?

 

 

This entire topic has nothing to do with gaming or consumer in any way. This is server/workstation stuff. The entire point of AMD's focus here, is to help programmers of CUDA based software, port their software/programming to a different non proprietary language and then compile it to something that will run really great on AMD hardware. The point is to gain market share in the CUDA based compute markets.

 

Oh, here was me getting excited...

[Notional]

Oh, here was me getting excited...

 

Be excited that CUDA gets competition. CUDA needs to die in the consumer market, completely. But Very few effects really utilizes APEX (CUDA) anyways, even in GameWorks, so meh.

[patrickjp93]

Be excited that CUDA gets competition. CUDA needs to die in the consumer market, completely. But Very few effects really utilizes APEX (CUDA) anyways, even in GameWorks, so meh.

Why does it need to die? If AMD can't offer better OpenCL performance and OpenCL is simply a total pain to program in, then why should Nvidia give up all that work? CUDA has been vastly superior for a decade. Enough with the butt hurt over Nvidia's success.

[Notional]

Why does it need to die? If AMD can't offer better OpenCL performance and OpenCL is simply a total pain to program in, then why should Nvidia give up all that work? CUDA has been vastly superior for a decade. Enough with the butt hurt over Nvidia's success.

The problem is proprietary tech. I would scold Intel and amd too for something like that.

Industry standards always results in better competition. Sure it's not in consumer space, but the issue still stands.

[patrickjp93]

The problem is proprietary tech. I would scold Intel and amd too for something like that.

Industry standards always results in better competition. Sure it's not in consumer space, but the issue still stands.

I'm sorry but that's not always true. CUDA has wiped the floor with OpenCL in performance even in open-source benchmarks like Linpack. When the best programmers in the world can't make AMD win under OpenCL despite having the theoretical advantage, it's not a coincidence. Heck even Intel with OpenMP as the recommended programming language for its Xeon Phi has done better at the industry benchmarks than AMD's best FirePro. Open standards do not always come out the winner, and I think we'll see that happen again with 3DXPoint which, on paper, blows Wide-IO and memristor tech clear out of the water.

Nvidia is no saint, but there is no justification in asking it to give up CUDA until a better alternative exists. OpenCL 2.1 has finally started catching up in the sensibility department (being able to use STL algorithms in-code without having to write them yourself), but it's still a fragmented programming language where kernels are completely separate from the main system code. Until someone comes up with a better alternative to CUDA, CUDA will reign in HPC.

[Dabombinable]

I'm sorry but that's not always true. CUDA has wiped the floor with OpenCL in performance even in open-source benchmarks like Linpack. When the best programmers in the world can't make AMD win under OpenCL despite having the theoretical advantage, it's not a coincidence. Heck even Intel with OpenMP as the recommended programming language for its Xeon Phi has done better at the industry benchmarks than AMD's best FirePro. Open standards do not always come out the winner, and I think we'll see that happen again with 3DXPoint which, on paper, blows Wide-IO and memristor tech clear out of the water.

Nvidia is no saint, but there is no justification in asking it to give up CUDA until a better alternative exists. OpenCL 2.1 has finally started catching up in the sensibility department (being able to use STL algorithms in-code without having to write them yourself), but it's still a fragmented programming language where kernels are completely separate from the main system code. Until someone comes up with a better alternative to CUDA, CUDA will reign in HPC.

At times what you say does come across as BS-but I agree with that^