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

[patrickjp93]

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

Called the Boltzmann Initiative, AMD is overhauling their entire approach to HPC, especially when it comes to GPGPU acceleration. Today AMD announced it will be implementing CUDA, OpenMP, C++ compilation for its GPUs.

Last but certainly not least in the Boltzmann Initiative is AMD’s effort to fully extend a bridge into the world of CUDA developers. With HCC to bring AMD’s programming environment more on par with what CUDA developers expect, AMD realizes that just being as good as NVIDIA won’t always be good enough, that developers accustomed to the syntax of CUDA won’t want to change, and that CUDA won’t be going anywhere anytime soon. The solution to that problem is the Heterogeneous-compute Interface for Portability, otherwise known as HIP, which gives CUDA developers the tools they need to easily move over to AMD GPUs.

The potential boost this brings to AMD's compute performance (as OpenCL is historically a fair deal worse than CUDA in the same workloads, be it for drivers or other reasons) is quite large, especially since theoretically AMD's FirePros have the compute advantage. At least now the cost of switching to AMD will be decreased for the HPC community: less libraries to rewrite, inter-architectural compatibility, etc..)

It seems AMD has finally come to its senses about programming standards that have proven more effective than OpenCL. The implications for this are actually great for AMD's gaming prospects too since so much of advanced PhysX is based in CUDA, giving AMD far better access than ever before to Nvidia's feature sets.

Being able to compile C++ code directly to the GPU will also save a ton in development time and skill ramp up for new developers.

Finally the red team sees reason... Maybe there is hope.

EDIT: while not stated in the article itself, I'm pretty sure the only way you can have a commercially used cross-compiler for CUDA is to have a CUDA license. It's a derived fact, not one provided in stone.

Edited November 18, 2015 by Godlygamer23
Edited title.

[PlayStation 2]

It's not that OpenCL was weak, it was that no one implemented it.

 

Either way, this isn't a good sign.

 

insert joke about nvidia not supporting opencl here

[AlwaysFSX]

Yay! We're unifying!

[don_svetlio]

It's a good thing - they are now supporting both.

[Peter Vaughan Truslow]

I don't think they're going to be abandoning OpenCL any time soon, this does give Devs more options. I'd beinterested to see how AMD cards stack up for CUDA performance, as it is well known that although they are comparable for graphics performance, AMD dominates Nvidia in terms of OpenCL performance.

I'm also intrigued about the C++ compilation, as C++ isn't natively multithreaded, and multithreading must be manually implented by the programmer.

[The_Stig]

I dont know exactly what his means, but I think its good.

[Khvarrioiren]

Took their sweet, sweet time that's for sure. Well, a step in the right direction, no doubt about that.

[patrickjp93]

It's not that OpenCL was weak, it was that no one implemented it.

 

Either way, this isn't a good sign.

 

insert joke about nvidia not supporting opencl here

Intel's Xeon Phi do, and they're far more popular than FirePros in the HPC world.

[Master Disaster]

To be honest, at this point Nvidia should just give AMD CUDA & PhysX licenses for free. It would only serve to improve everyone's experience by unifying the technology across all devices.

[Starelementpoke]

I dont know exactly what his means, but I think its good.

Same.

[patrickjp93]

I don't think they're going to be abandoning OpenCL any time soon, this does give Devs more options. I'd beinterested to see how AMD cards stack up for CUDA performance, as it is well known that although they are comparable for graphics performance, AMD dominates Nvidia in terms of OpenCL performance.

I'm also intrigued about the C++ compilation, as C++ isn't natively multithreaded, and multithreading must be manually implented by the programmer.

Where have you been? Native multithreading's been here since C++11.

 

Compiler flag -D_GLIBCXX_PARALLEL https://gcc.gnu.org/onlinedocs/libstdc++/manual/parallel_mode.html

[don_svetlio]

To be honest, at this point Nvidia should just give AMD CUDA & PhysX licenses for free. It would only serve to improve everyone's experience by unifying the technology across all devices.

But then they'd have to optimize their blackbox libraries - god forbid that!

[patrickjp93]

To be honest, at this point Nvidia should just give AMD CUDA & PhysX licenses for free. It would only serve to improve everyone's experience by unifying the technology across all devices.

Nvidia offered CUDA for free several times. AMD turned its nose up at supporting a proprietary standard. Now it seems finally someone has a lick of sense.

 

Source: http://www.extremetech.com/computing/82264-why-wont-ati-support-cuda-and-physx

 

AMD/ATI's problem has always been the "Nvidia isn't very collaborative" argument, and frankly that's BS to me.

[patrickjp93]

But then they'd have to optimize their blackbox libraries - god forbid that!

They aren't black box. Anyone with a brain and llvm can crack them wide open via disassembly and decompiling.

[don_svetlio]

Nvidia offered CUDA for free several times. AMD turned its nose up at supporting a proprietary standard. Now it seems finally someone has a lick of sense.

I love you too honey

[don_svetlio]

They aren't black box. Anyone with a brain and llvm can crack them wide open via disassembly and decompiling.

That's Illegal hun

[patrickjp93]

That's Illegal hun

No it's not. It's illegal to publish that code. Someone needs to read the licenses more carefully.

[Master Disaster]

Didn't Nvidia already partially open source Gameworks or was I imagining things? I thought I'd read that they no longer forbid devs from passing on Game Works source code and were allowing AMD access to the code to prove they were not sabotaging AMD devices.

[patrickjp93]

Didn't Nvidia already partially open source Gameworks or was I imagining things? I thought I'd read that they no longer forbid devs from passing on Game Works source code and were allowing AMD access to the code to prove they were not sabotaging AMD devices.

They open-sourced CPU-based PhysX, not GPU-based.

[don_svetlio]

Didn't Nvidia already partially open source Gameworks or was I imagining things? I thought I'd read that they no longer forbid devs from passing on Game Works source code and were allowing AMD access to the code to prove they were not sabotaging AMD devices.

If devs pay, yes.

[Fictionvl]

This is really good for us consumers, more competition = lower prices!

[zMeul]

this may not have any implication on the PC gaming market unless AMD goes the full length and gets GPU Physx for the Radeon

as is stands, AMD looks really interested in the HPC segment

[patrickjp93]

this may not have any implication on the PC gaming market unless AMD goes the full length and gets GPU Physx for the Radeon

as is stands, AMD looks really interested in the HPC segment

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

[LAwLz]

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:

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.

 

CUDA is NVIDIA’s, through and through, and it does make one wonder whether NVIDIA would try to sue AMD for implementing the CUDA API without NVIDIA’s permission, particularly in light of the latest developments in the Oracle vs. Google case on the Java API.

 

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.

[Levent]

interesting, I wonder if any of the current and previous gpu lineups going to support it