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

[Misanthrope]

Cuda translation as it was called elsewhere probably means overhead so the lead firepro cards have might bring them right back down to being "equal" by spending a lot more muscle to do so...not unlike gaming in fact.

[LAwLz]

@MageTank

 

Patrick posted source...

 

patrick. i am so proud of you

Too bad his own source disagrees with him this time.

 

 

I think it is jumping to conclusions to say that AMD got a CUDA license. Making your own compiler accept code from one language does not, as far as I know, require a license as long as the output is not the same. I mean, it's not like Apple and Google gave Microsoft a license for their compatibility bridges (or whatever they call it) used in the iOS/Android -> Metro app compiler.

I doubt Microsoft gave Sony a license so that developers could reuse code from their Xbone games for the PS4 versions either.

 

 

 

Cuda translation as it was called elsewhere probably means overhead so the lead firepro cards have might bring them right back down to being "equal" by spending a lot more muscle to do so...not unlike gaming in fact.

It might not necessary mean more overhead. It's not emulation after all.

There might be some performance penalties but that remains to be seen. It's not guaranteed though.

[Prysin]

 

Cuda translation as it was called elsewhere probably means overhead so the lead firepro cards have might bring them right back down to being "equal" by spending a lot more muscle to do so...not unlike gaming in fact.

given the stark overdimensioned specs of the W9100 compared to say a K6000.... well... even then it will have a HUGE +1 on Nvidia if it can even remotely run CUDA code.

 

If Firepro can run CUDA, and perform equal to their Nvidia counterpart they will be selling like hotcakes. Simply because they are a magnitude of order cheaper (Firepro W9100 is 1000 USD cheaper then the Quadro K6000)

 

 

Too bad his own source disagrees with him this time.

 

I think it is jumping to conclusions to say that AMD got a CUDA license. Making your own compiler accept code from one language does not, as far as I know, require a license as long as the output is not the same. I mean, it's not like Apple and Google gave Microsoft a license for their compatibility bridges (or whatever they call it) used in the iOS/Android -> Metro app compiler.
I doubt Microsoft gave Sony a license so that developers could reuse code from their Xbone games for the PS4 versions either.

dude, he tried... its like the first time he did this without 12 pages of text coaxing him into doing it.

 

 

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

don't be mad...

 

i ain't... been trolled so many times while roaming on /b/ that i started to like the song by now.

[Nowak]

Well, this is an interesting development.

[patrickjp93]

it wouldnt be the first company to try that

 

*cough*Nvidia Tegra*cough*

Nvidia was in a financial position to take that hit and was smart enough to back down in an instant. AMD can't survive much more bad PR, and it certainly can't afford to be fighting off a class-action lawsuit on one side and then get hammered by Nvidia's legal team.

[patrickjp93]

Cuda translation as it was called elsewhere probably means overhead so the lead firepro cards have might bring them right back down to being "equal" by spending a lot more muscle to do so...not unlike gaming in fact.

You assume the translation will be happening in real time instead of at compile time (the smarter way to do it).

 

Too bad his own source disagrees with him this time.

 

 

I think it is jumping to conclusions to say that AMD got a CUDA license. Making your own compiler accept code from one language does not, as far as I know, require a license as long as the output is not the same. I mean, it's not like Apple and Google gave Microsoft a license for their compatibility bridges (or whatever they call it) used in the iOS/Android -> Metro app compiler.

I doubt Microsoft gave Sony a license so that developers could reuse code from their Xbone games for the PS4 versions either.

 

 

 

It might not necessary mean more overhead. It's not emulation after all.

There might be some performance penalties but that remains to be seen. It's not guaranteed though.

It doesn't disagree with me at all. Having a license doesn't mean AMD's implementing it in hardware.

 

Please see the CUDA SDK EULA. A Cross-Compiler will require a licensing agreement. http://docs.nvidia.com/cuda/eula/index.html#abstract 

 

Google's GO is the same way.

 

Apple and Google have a vested interest in letting Microsoft run Apple and Google apps. There's no reason to fight their efforts on it. Straw man argument, your typical fallacy.

 

Microsoft didn't have to. The code base belonged to the development studios. Unless by contract (or an EULA) with Microsoft they forbade it, Microsoft would have no leverage to dictate it. Again, Straw Man.

[Prysin]

Nvidia was in a financial position to take that hit and was smart enough to back down in an instant. AMD can't survive much more bad PR, and it certainly can't afford to be fighting off a class-action lawsuit on one side and then get hammered by Nvidia's legal team.

doesnt matter mate...

 

look at scale..

Nvidia vs Intel

Nvidia vs AMD

 

it doesnt matter, in either case the owner of the IP can bury the other company. Intel could sue for much higher damages because their market share alone would allow for "much more affected clients" % wise. Same goes for Nvidia... they have the lions share of HPC market (when comparing AMD to Nvidia). So their potential for loss is much greater then just a petty sum alone.

 

Also, there is a final thing though.

neither you, nor me knows the exact extent of the CUDA IP..... if Nvidia didnt cover their asses good enough, AMD might have found a loophole, which is why they did this.

[patrickjp93]

doesnt matter mate...

 

look at scale..

Nvidia vs Intel

Nvidia vs AMD

 

it doesnt matter, in either case the owner of the IP can bury the other company. Intel could sue for much higher damages because their market share alone would allow for "much more affected clients" % wise. Same goes for Nvidia... they have the lions share of HPC market (when comparing AMD to Nvidia). So their potential for loss is much greater then just a petty sum alone.

 

Also, there is a final thing though.

neither you, nor me knows the exact extent of the CUDA IP..... if Nvidia didnt cover their asses good enough, AMD might have found a loophole, which is why they did this.

10 million in legal fees is very different for Nvidia and AMD. 

 

Nvidia leaving a legal loophole? Laughable. Are we forgetting the Havok vs. PhysX lawsuit of 2009? Nvidia is air-tight on IP security.

[Prysin]

10 million in legal fees is very different for Nvidia and AMD. 

 

Nvidia leaving a legal loophole? Laughable. Are we forgetting the Havok vs. PhysX lawsuit of 2009? Nvidia is air-tight on IP security.

when was CUDA licensed?

when was PhysX licensed?

[Misanthrope]

You assume the translation will be happening in real time instead of at compile time (the smarter way to do it).

 

Fair enough. All the better means Nvidia might actually have to revise costs and such.

[LAwLz]

It doesn't disagree with me at all. Having a license doesn't mean AMD's implementing it in hardware.

Please see the CUDA SDK EULA. A Cross-Compiler will require a licensing agreement. http://docs.nvidia.com/cuda/eula/index.html#abstract

Google's GO is the same way.

Apple and Google have a vested interest in letting Microsoft run Apple and Google apps. There's no reason to fight their efforts on it. Straw man argument, your typical fallacy.

Microsoft didn't have to. The code base belonged to the development studios. Unless by contract (or an EULA) with Microsoft they forbade it, Microsoft would have no leverage to dictate it. Again, Straw Man.

But it does disagree with you. You are saying they got a license for it while your source are wondering if Nvidia might sue AMD for doing this.

The EULA does not mention anything about cross-compiling. All it says is that if you are going to compile the SDK from source then you are not allowed to do X, Y and Z. All other mentions of compiling is just "you're not allowed to decompile the SKD". I can't see anything about Nvidia owning the rights to any code written for CUDA. However, I have not read through the entire document so if you have found a line saying that then feel free to quote it here.

 

I won't comment on Google GO since I don't know enough about it.

 

Ehm... Why would Apple and Google want Microsoft to run programs from their platforms on Windows?

[Prysin]

But it does disagree with you. You are saying they got a license for it while your source are wondering if Nvidia might sue AMD for doing this.

The EULA does not mention anything about cross-compiling. All it says is that if you are going to compile the SDK from source then you are not allowed to do X, Y and Z. All other mentions of compiling is just "you're not allowed to decompile the SKD". I can't see anything about Nvidia owning the rights to any code written for CUDA. However, I have not read through the entire document so if you have found a line saying that then feel free to quote it here.

 

I won't comment on Google GO since I don't know enough about it.

 

Ehm... Why would Apple and Google want Microsoft to run programs from their platforms on Windows?

cuz apple/google would make money off those apps....

you know, most of the stuff you publish in the app store(s) falls under some distorted version of the good old "whatever you upload here, we can do whatever we fuck with it" term.

[deviant88]

All i know is that blender users will be able to probably use AMD card for Cycles rendering with CUDA.

[patrickjp93]

when was CUDA licensed?

when was PhysX licensed?

As for CUDA, clearly before AMD presented this. These deals aren't exactly done in public spotlight.

PhysX, never, hence why Nvidia sued Intel when the Havoc engine started using the same render pipelines as PhysX circa 2009.

[RagnarokDel]

Does it mean that AMD cards could suck less at nvidia gameworks in the future?

[patrickjp93]

But it does disagree with you. You are saying they got a license for it while your source are wondering if Nvidia might sue AMD for doing this.

The EULA does not mention anything about cross-compiling. All it says is that if you are going to compile the SDK from source then you are not allowed to do X, Y and Z. All other mentions of compiling is just "you're not allowed to decompile the SKD". I can't see anything about Nvidia owning the rights to any code written for CUDA. However, I have not read through the entire document so if you have found a line saying that then feel free to quote it here.

 

I won't comment on Google GO since I don't know enough about it.

 

Ehm... Why would Apple and Google want Microsoft to run programs from their platforms on Windows?

Because the source doesn't understand the obvious ramifications.

 

Are you seriously not able to extrapolate. Compiling from source, including cross-compiling, is covered by that. AMD cannot have a commercially functioning cross-compiler without a license to do so, as outlined by the EULA of the CUDA dev kit, and no one is allowed the use of such an unlicensed cross-compiler by the same. It's crystal clear to me, and I'm not even a lawyer.

 

It's not that Nvidia owns the code you write for CUDA. It's that it owns the usage thereof.

 

Money, duh.

[patrickjp93]

Does it mean that AMD cards could suck less at nvidia gameworks in the future?

Potentially yes.

[Guest]

You know, to all the people saying AMD made OpenCL, no they didn't.

They just know how to implement the OpenCL Technology better than NVidia

On OP:

Yay, more proprietary standards dominating the Open Source version.

[Prysin]

 

Potentially yes.

probably not...

 

geometry capabilities of the GCN GPU architecture is the bit that hurts the most. Not all the coding in failworks.
you know, if failworks was actually CODED to screw over AMD. Nvidia would be sued into the abyss for anti competitive behaviour. There would be nothing left of them, at all.

 

so instead, they just use lots of effects AMDs GCN architecture doesnt handle very well, but maxwell does. Simply put, they designed their own stuff to excel where the competition did not, then delivered the final blow by adapting games to use these effects more aggressively.

 

 

You know, to all the people saying AMD made OpenCL, no they didn't.

 

They just know how to implement the OpenCL Technology better than NVidia

 

On OP:
Yay, more proprietary standards dominating the Open Source version.

OpenCL is Khronos group, everyone knows that

[ACatWithThumbs]

This is amazing I was hoping for years something like this would happen.
Not only will it improve CUDA but we might see PC based games have special CUDA features.
And this should also help with PC optimization in general as it gives a unified compute platform for PC's.

 

[patrickjp93]

You know, to all the people saying AMD made OpenCL, no they didn't.

They just know how to implement the OpenCL Technology better than NVidia

On OP:

Yay, more proprietary standards dominating the Open Source version.

BS. Nvidia just doesn't focus on it because has CUDA and a fairly good ecosystem and marketshare. OpenCL 2.0 support is inbound.

 

Nvidia is the king of compute, and arguably CUDA is way ahead on compute technologies than OpenCL is. Asynchronous compute kernels was in CUDA long before it was in OpenCL (note this is different from doing graphics rendering and computing kernels at the same time asynchronously).

 

 

OpenCL is Khronos group, everyone knows that

 

Technically it's Apple specifically.

 

 

This is amazing I was hoping for years something like this would happen.

Not only will it improve CUDA but we might see PC based games have special CUDA features.

And this should also help with PC optimization in general as it gives a unified compute platform for PC's.

 

If you wanted a unified compute platform, Intel developed it back in 2000 with OpenMP.

[LAwLz]

Because the source doesn't understand the obvious ramifications.

 

Are you seriously not able to extrapolate. Compiling from source, including cross-compiling, is covered by that. AMD cannot have a commercially functioning cross-compiler without a license to do so, as outlined by the EULA of the CUDA dev kit, and no one is allowed the use of such an unlicensed cross-compiler by the same. It's crystal clear to me, and I'm not even a lawyer.

 

It's not that Nvidia owns the code you write for CUDA. It's that it owns the usage thereof.

The "compile from source" Nvidia is talking about is compiling the SDK, not every single program written for CUDA. There is a huge difference. AMD is not allowed to compile and redistribute the CUDA SDK, but I don't see anything hindering them from accepting CUDA source code and compiling that to some other language. Can you please link me to something that says you are not allowed to compile CUDA coda you have written to another language? I can't find it.

I don't want some "it is implied" or anything like that either. User Agreements are not written in language which requires you to guess things. If Nvidia says you are not allowed to compile and redistribute the SDK then they are talking about the SDK, not the SDK and every other program you write for CUDA.

 

 

 

Money, duh.

cuz apple/google would make money off those apps....

you know, most of the stuff you publish in the app store(s) falls under some distorted version of the good old "whatever you upload here, we can do whatever we fuck with it" term.

Apple and Google would not make money off the apps because the converted apps will be sold through Microsoft's store. The ad service is also changed out for Microsoft's so Apple and Google will not see a single cent from the cross-compiled apps.

[Prysin]

The "compile from source" Nvidia is talking about is compiling the SDK, not every single program written for CUDA. There is a huge difference. AMD is not allowed to compile and redistribute the CUDA SDK, but I don't see anything hindering them from accepting CUDA source code and compiling that to some other language. Can you please link me to something that says you are not allowed to compile CUDA coda you have written to another language? I can't find it.

I don't want some "it is implied" or anything like that either. User Agreements are not written in language which requires you to guess things. If Nvidia says you are not allowed to compile and redistribute the SDK then they are talking about the SDK, not the SDK and every other program you write for CUDA.

 

 

 

Apple and Google would not make money off the apps because the converted apps will be sold through Microsoft's store. The ad service is also changed out for Microsoft's so Apple and Google will not see a single cent from the cross-compiled apps.

royalties. If there is a sale and the original distributor is X, then X has claim to royalties unless the creator itself has formed a contract with the second party to sell via their stores.

But since we are talking about porting apps over, then royalties for sure.

 

Everything is about money. If apple and google didnt make any form of revenue out of this ordeal they would sue MS a long time ago

[SurvivorNVL]

Oh, this is exciting.  So, this will likely come to consumer cards after Arctic Islands, but this is a great change.  Would be useful for their APU's as well.

[patrickjp93]

The "compile from source" Nvidia is talking about is compiling the SDK, not every single program written for CUDA. There is a huge difference. AMD is not allowed to compile and redistribute the CUDA SDK, but I don't see anything hindering them from accepting CUDA source code and compiling that to some other language. Can you please link me to something that says you are not allowed to compile CUDA coda you have written to another language? I can't find it.

I don't want some "it is implied" or anything like that either. User Agreements are not written in language which requires you to guess things. If Nvidia says you are not allowed to compile and redistribute the SDK then they are talking about the SDK, not the SDK and every other program you write for CUDA.

 

 

 

Apple and Google would not make money off the apps because the converted apps will be sold through Microsoft's store. The ad service is also changed out for Microsoft's so Apple and Google will not see a single cent from the cross-compiled apps.

1) It applies to all the code. My university actually discussed this with Nvidia representatives since we're getting a Xeon Phi node.

 

2) Google and Apple both make money. They license those apps.