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

[zMeul]

GPU PhysX is purely CUDA code. They don't need a separate license to be able to support it.

yes, but nVidia doesn't allow access to GPU based PhysX - nVidia has to give them access

 

there are architectural differences, AMD need to compile their own PhysX library

[Sir Asvald]

So would the current cards get the update via the driver? Or they will implement them on the newer cards?

[Notional]

WTF Patrick, has your reading comprehension completely dissapered? AMD is not adopting CUDA what so ever. They have created a new compiler that introduces HSA to GPGPU's and makes everything easier for deveoplers. As for CUDA, they have introduced a syntax that makes it easier for CUDA programmers to port/code for HIP (AMD's new compiler)

 

 

To be clear here, HIP is not a means for AMD GPUs to run compiled CUDA programs. CUDA is and remains an NVIDIA technology. But HIP is the means for source-to-source translation, so that developers will have a far easier time targeting AMD GPUs. Given that the HPC market is one where developers are typically writing all of their own code here anyhow and tweaking it for the specific architecture it’s meant to run on, a source-to-source translation covers most of AMD’s needs right there, and retains AMD’s ability to compile CUDA code from a high level where they can better optimize that code for their GPUs.​

 

[Trixanity]

From reading the article, I got the sense that AMD does not and will never have a CUDA license and it's still an Nvidia-only technology. From my understanding, AMD have developed tools to "translate" CUDA to whatever AMD uses, so CUDA developers can more easily develop with AMD solutions in mind as well. So effectively AMD still very much supports OpenCL but kinda also CUDA as long as developers embrace it, as CUDA in its native form does in fact not work on AMD cards, not even with these changes. So developers will still need to get on board.

[patrickjp93]

Did they really get a CUDA license? The way I read it, it's just a compiler that will accept CUDA code (i.e. it ports CUDA code to something else). On top of that, it says "simplify porting from CUDA to HIP" which to me means "developers will still have to rework their CUDA code to make it work on AMD".

I don't think you need a CUDA license to do this.

Edit (after reading the announcement):

The announcement does not mention AMD getting a license either. The article on Anandtech even clarifies this by saying:

Pretty cool news though. I don't think it will have a big effect on the consumer market but it might help AMD in the HPC market.

The only way to cross-compile CUDA programs afaik is to have a CUDA license.

[patrickjp93]

So would the current cards get the update via the driver? Or they will implement them on the newer cards?

By the time this is fully implemented, you'll be on a new card anyway.

[patrickjp93]

WTF Patrick, has your reading comprehension completely dissapered? AMD is not adopting CUDA what so ever. They have created a new compiler that introduces HSA to GPGPU's and makes everything easier for deveoplers. As for CUDA, they have introduced a syntax that makes it easier for CUDA programmers to port/code for HIP (AMD's new compiler)

Not at all. Much like Google's GO language, you can't make a compiler or cross-compiler without a license, and that's bi-directional. The only way for AMD to be able to do this is to have a license.

Is this tiny mental leap really so hard for you people? I didn't jump to my other conclusion which was that AMD has an MIC accelerator incoming, as it's the only reason to put this much effort into ubiquity.

[don_svetlio]

Not at all. Much like Google's GO language, you can't make a compiler or cross-compiler without a license, and that's no-directional. The only way for AMD to be able to do this is to have a license.

Look above patrick - your own source says otherwise

[patrickjp93]

Look above patrick - your own source says otherwise

No it doesn't. What the source says is that AMD won't be running CUDA-compiled code. It's a cross-compiler (likely to SPIR-V). You still need a license to do that. Nowhere in the source article is this contradicted in any way.

[Maslofski]

LOOOL finally

[Guest]

I highly doubt they will abandon openCL. 

[The Benjamins]

No it doesn't. What the source says is that AMD won't be running CUDA-compiled code. It's a cross-compiler (likely to SPIR-V). You still need a license to do that. Nowhere in the source article is this contradicted in any way.

 

If they have a license why would people be scaried Nvidia sueing AMD of this?

[patrickjp93]

If they have a license why would people be scaried Nvidia sueing AMD of this?

Because people didn't use their heads in the first place. AMD wouldn't be announcing CUDA compatibility unless it had already been worked out. Just because average humans are overall irrational and don't think ahead doesn't mean I'm bound to the average. AMD's engineers certainly aren't average, though their business leaders have certainly been mediocre.

 

With a license in-hand, AMD doesn't have to rework its architectures and instruction sets. It just has to make a translation software to go from Nvidia's code standard to whatever AMD wants, which as I mentioned is likely the SPIR-V instruction set since that's what Vulkan/GLNext is based on.

[Prysin]

http://www.anandtech.com/show/9792/amd-sc15-boltzmann-initiative-announced-c-and-cuda-compilers-for-amd-gpus

 

@MageTank

 

Patrick posted source...

 

patrick. i am so proud of you

[patrickjp93]

@MageTank

 

Patrick posted source...

 

patrick. i am so proud of you

I always do. Source

[Prysin]

I always do. Source

don't push your luck

[MageTank]

I always do. Source

Can't believe i fell for that. Oh well. Two can play at this game.

[patrickjp93]

Can't believe i fell for that. Oh well. Two can play at this game.

Don't take it too hard man. Take Five and Chill

[Sir Asvald]

By the time this is fully implemented, you'll be on a new card anyway.

I like my 7950.

[SilicateWielder]

Can we expect the APUs coming out next year to support this? Or will it be another year or two?

[patrickjp93]

Can we expect the APUs coming out next year to support this? Or will it be another year or two?

It all depends on how fast AMD can get this compiler up and running. Starting with LLVM was good, but it will take time.

[SilicateWielder]

It all depends on how fast AMD can get this compiler up and running. Starting with LLVM was good, but it will take time.

 

Ah, okay. 

[Prysin]

Not at all. Much like Google's GO language, you can't make a compiler or cross-compiler without a license, and that's bi-directional. The only way for AMD to be able to do this is to have a license.

Is this tiny mental leap really so hard for you people? I didn't jump to my other conclusion which was that AMD has an MIC accelerator incoming, as it's the only reason to put this much effort into ubiquity.

no... AMD can make this and hope Nvidia doesnt see them as enough of a threat (just yet) to sue the shit out of them for basically dancing around the outskirts of their IPs

[patrickjp93]

no... AMD can make this and hope Nvidia doesnt see them as enough of a threat (just yet) to sue the shit out of them for basically dancing around the outskirts of their IPs

AMD may take risks, but it cannot possibly be this stupid.

[Prysin]

AMD may take risks, but it cannot possibly be this stupid.

it wouldnt be the first company to try that

 

*cough*Nvidia Tegra*cough*