Cuda's future

[Arkadrel]

I may be wrong but I think nvidia is betting too high for CUDA.

Microsoft recently released their first server OS for HPC. They released DirectCompute in the last two releases of the DX spec, I doubt that's only for games. I guess they want to compete with Linux/Unix in the HPC front as otherwise the wouldn't release an OS for it and that involves providing a good performing, powerful and familiar API (specially the 'familiar' part is what Microsoft is good at) for GPGPUs that is vendor agnostic.

At some point in time they nvidia and MS will be competing on software. MS will be able to offer an API for AMD and nvidia (and Intel?) cards while nvidia will offer an API for their cards. Who will win?

Use the same reasoning for OpenCL and Linux/Unix(/Mac) world.

Add to this that, at some point in time, AMD will start to take GPGPU more seriously (probably when their OpenCL API starts getting mature enough). And Intel. Add again APUs.

How much will CUDA pay off in the next 2, 5 and 10 years? I see diminishing returns here, while their hardware will be compromised size-wise (and more expensive in the consumer market, who hardly cares anything about GPGPU performance) just to support a market that they currently dominate, but that will be quite more contested in a couple of years. A software API is a long-term investment, and in this case their ahead start won't be worth anything in 2015 even if they are alive by then. CUDA is something cool to give a nice income of cash in the PRO and HPC market for a few years with high margin parts but nothing more, without forgetting that their primary target market is consumer electronics.

I see a black screen of death coming in for the CUDA based OS running on Fermi... /sarcasm

About why is this relevant to this topic and not completely unrelated, I think there is a relation between CUDA being pushed hard by nvidia and their architectural problems in the consumer front and that if they were more in touch with the reality of their primary target market they would be more successful. Please let me know if this post doesn't make any sense, I'm having trouble trying to make my points clear recently

From S/A forums.

Is Mircrosoft makeing a differnt version of cuda that works with intel/amd/nvidia (all) cards?

Others are saying the most common use for GPGPU is encodeing... and that Intel new CPUs will have a unit that does that and beats the GPUs at it.

Is Cuda destined to fail? What do you guys think on this matter?

 

[formulav8]

If CUDA does fail you can probably pinpoint non-open API as the reason.

 

[Arkadrel]

Mircosoft are kinda a big deal in the software world.. and "A software API is a long-term investment" plus the fact that a general API that can work on all cards seems more likely to get supported.

Yeah that fact that Cuda is nvidia only might have doomed it, and going toe to toe with mircosoft is gonna be tough reguardless. Also with Intel stealing one of the features that benefits the avg. joe consumer from cuda (faster encodeing/decodeing) with their cpus.

I dont think GPGPU will go away (in nvidia hardware). I just suspect a few years from now even nvidia users will be useing the mircosoft API, so it wont really be a selling point for them anymore (and it seems like they put alot into developing it, must suck that MS decided to give them competition).

If CUDA does fail you can probably pinpoint non-open API as the reason.

Do you suspect the same will happend to Physx? Because that hasnt really taken off either.

 

[Genx87]

I believe Microsofts version is called DirectCompute and is a big part of DX11. Nvidia does just fine using DirectCompute compared to the competition. But you have to remember in the market this is targeted at Nvidia holds 90% marketshare. So right now Cuda is the choice of developers until AMD gets off its ass and delivers a product that can compete.

PhysX will most likely always be Nvidia's physics for their cards. But you may see something like OpenCL being used to make the calls to their api. Nvidia supports OpenCL just fine. There were some benchmarks about a month ago where the bullet physics API using OpenCL looked fantastic on Nvidia. Where it failed to run on some of AMDs hardware.

Edit: That guy seems to be under some misunderstanding OpenCL is AMD. Nvidia supports it and has done a better job at providing driver and compatibility and performance than AMD. And it is interesting he discounts Microsofts monopoly and position that DirectCompute only works within a microsoft OS. A lot of the target of these programs that Cuda or OpenCL will eventually be will not be on a microsoft OS. Pretty sure that big super computer over in china that is using tesla is not running Server 2008 Where Cuda is platform agnostic. It should work in a Microsoft, Apple OSX, or Linux envrionment.

 

[cbn]

Also with Intel stealing one of the features that benefits the avg. joe consumer from cuda (faster encodeing/decodeing) with their cpus.

Is Intel's new Sandy Bridge decoder really fast enough to threaten CUDA or OpenCL?

EDIT: Nvm, I found this--->http://www.techradar.com/news/compu...ntel-chips-encode-hd-video-in-seconds--716246 (Apparently specific modifications to Windows 7 and third party software need to be made to gain the speed-up benefits)

 

[n0x1ous]

From S/A forums.

Is Mircrosoft makeing a differnt version of cuda that works with intel/amd/nvidia (all) cards?

Others are saying the most common use for GPGPU is encodeing... and that Intel new CPUs will have a unit that does that and beats the GPUs at it.

Is Cuda destined to fail? What do you guys think on this matter?

Ah yes, more nvidia is doomed conjecture.

In reality, its more like nvidia keeps on actively developing useful stuff and AMD and its fan's keep on whining about how its not open source, won't be accepted yada yada......open source is great but when its proprietary vs nothing its not really a problem.

IF cuda does get eventually supplanted, nvidia won't really be that upset. They support all the open standards better then AMD NOW anyway. they developed their own standard to get the ball rolling unlike AMD.

 

[Keysplayr]

Ah yes, more nvidia is doomed conjecture.

In reality, its more like nvidia keeps on actively developing useful stuff and AMD and its fan's keep on whining about how its not open source, won't be accepted yada yada......open source is great but when its proprietary vs nothing its not really a problem.

IF cuda does get eventually supplanted, nvidia won't really be that upset. They support all the open standards better then AMD NOW anyway. they developed their own standard to get the ball rolling unlike AMD.

Well said. I wonder how long AMD fans will put up with that. How long before they reach up and scratch their heads saying, "hmmmm.".

 

[Arkadrel]

Still makeing a product that competes with Mircosoft in software.
That has the support of Intel and AMD...

Anyone here remember 3dfx and GLIDE? that went up against mircosofts opengl and Dx.

Where is the Glide API today? and where are 3dfx?

 

[Keysplayr]

Still makeing a product that competes with Mircosoft in software.
That has the support of Intel and AMD...

Anyone here remember 3dfx and GLIDE? that went up against mircosofts opengl and Dx.

Where is the Glide API today? and where are 3dfx?

"that went up against mircosofts opengl and Dx."
...and ATI, and Nvidia.

Nvidia will support any standard MS can dish out along with it's own CUDA which in it's current incarnation, is far beyond OpenCL, DirectCompute levels.

CUDA is going to do just fine. Even if it just remains to the professional level market. It isn't going to fail. Too far into it now.

 

[Genx87]

Still makeing a product that competes with Mircosoft in software.
That has the support of Intel and AMD...

Anyone here remember 3dfx and GLIDE? that went up against mircosofts opengl and Dx.

Where is the Glide API today? and where are 3dfx?

3DFX didnt fail because of Glide. It failed because it purchased STB. And overnight dropped all of its vendors pushing them to Nvidia and ATI and cut off a majority of the market to their hardware. I remember the day when the news came out about them purchasing STB. It was only a matter of time until they collapsed.

That said Nvidia wont have a problem supporting directcompute in the future. And how much support does OpenCL really have from Intel and AMD? AMD cant even be bothered to get the Physics API they support(Bullet) to work on their hardware using OpenCL right now.

 

[Scali]

nVidia supports all three options: C++ for Cuda, DirectCompute and OpenCL.
The advantage of Cuda is currently that it is more advanced and more powerful than the other two.
DirectCompute and OpenCL suffer from the least-common-denominator model, and are at about the level of the earliest Cuda, designed for G80. A very simple C-like language, much like HLSL/GLSL.
Cuda has since evolved considerably, and supports full C++ with virtualized memory, function pointers and recursion.
It may take years until DirectCompute and OpenCL get to this point. For HPC, it is not important that code is hardware-dependent anyway. It's not like conventional supercomputers are compatible with eachother on the software level.

 

[Schadenfroh]

Where is the Glide API today? and where are 3dfx?

3dfx died because they cut all of their AIB makers off when they purchased STB (main reason), focused too much on rehashed short term stop-gap products and not on rampage, in addition to spending like madmen on frivolous things like extravagant meals / parties till the day they closed their doors.

Glide was a plus for 3dfx, in a similar fashion, EAX support is really the only reason to consider creative. Proprietary APIs do not cause a company to fail if they support other widely used APIs well in addition to their proprietary ones.

 

[Seero]

From S/A forums.

Is Mircrosoft makeing a differnt version of cuda that works with intel/amd/nvidia (all) cards?

Others are saying the most common use for GPGPU is encodeing... and that Intel new CPUs will have a unit that does that and beats the GPUs at it.

Is Cuda destined to fail? What do you guys think on this matter?

In case you don't yet know, Nvidia supports DirectCompute by having DirectCompute sitting on top of their CUDA architecture. What is interesting here is not only Nvidia support DirectCompute, but also support their proprietary CUDA based FPs(Nvidia called it "Nvidia's CUDA extensions").

It seems that Nvidia's proprietary CUDA isn't your cup of tea, but Microsoft's proprietary DirectX is. While you may believe that DirectX is more open compare to CUDA because AMD can support it, it actually won't run on
"Android" or "ios".

 

[Scali]

It seems that Nvidia's proprietary CUDA isn't your cup of tea, but Intel's proprietary DirectX is.

I think you meant Microsoft there, not Intel.

 

[Seero]

I think you meant Microsoft there, not Intel.

Fixed.

 

[Cogman]

Cuda can DIAF for all I care. I don't like software techs that are tied to one companies hardware. I'm hoping that OpenCL or ever DirectCompute take off instead.

 

[Dark Shroud]

While you may believe that DirectX is more open compare to CUDA because AMD can support it, it actually won't run on
"Android" or "ios".

I'm going to take a wild guess here and bet that Cuda also doesn't work on iOS or Android.

In all seriousnes who cares. A program that uses Direct Compute, Open CL, or even Cuda most likly will not be used on a Mobile OS.

 

[nenforcer]

CUDA will remain relevant as long as its as closely tied to the hardware as it is.

You may have a neutral platform like OpenCL and DirectCompute which works on either AMD / Intel / nVidia GPU's but its performance and robustness may not match something like CUDA which is so tightly coupled.

At least nVidia I believe is banking on this!

 

[Genx87]

Cuda can DIAF for all I care. I don't like software techs that are tied to one companies hardware. I'm hoping that OpenCL or ever DirectCompute take off instead.

So you dont like software that is tied to one companies hardware(Nvidia) but are ok with software that is tied to one companies software(Microsoft DirectCompute)? Makes complete and logical sense.

 

[Genx87]

I'm going to take a wild guess here and bet that Cuda also doesn't work on iOS or Android.

In all seriousnes who cares. A program that uses Direct Compute, Open CL, or even Cuda most likly will not be used on a Mobile OS.

I dont see why Cuda couldnt work on those given an Nvidia chipset with approriate drivers. I would be swfully surprised if Cuda doesnt filter down to the Tegra line eventually.

 

[Scali]

I dont see why Cuda couldnt work on those given an Nvidia chipset with approriate drivers. I would be swfully surprised if Cuda doesnt filter down to the Tegra line eventually.

Yea, Cuda is multiplatform, it already works on linux and MacOS, close cousins of Android and iOS.

 

[Cogman]

So you dont like software that is tied to one companies hardware(Nvidia) but are ok with software that is tied to one companies software(Microsoft DirectCompute)? Makes complete and logical sense.

Actually, it does. I don't mind C# (initially only microsoft), what I mind is a language that can only be used on AMD CPUs or Intel CPUs. Software is relatively easier to port over than hardware is.

 

[Genx87]

Actually, it does. I don't mind C# (initially only microsoft), what I mind is a language that can only be used on AMD CPUs or Intel CPUs. Software is relatively easier to port over than hardware is.

You could port Cuda to OpenCL or DirectCompute.

 

[Cogman]

You could port Cuda to OpenCL or DirectCompute.

Sort of... Emulation and straight forward software conversions rarely works.

 

[Seero]

Actually, it does. I don't mind C# (initially only microsoft), what I mind is a language that can only be used on AMD CPUs or Intel CPUs. Software is relatively easier to port over than hardware is.

You forgot to mention Mac CPU. I still remember the days where people use floppy disks, and floppy that contain data from apple is not readable by Microsoft, and vice versa.

Many games run on PC, but not Mac. Where have you been all these years?

This isn't about whether CUDA should go open or not. This is about how OP believes that DirectCompute is better than CUDA. OpenGL is not new. In fact, the open source code idea has been there for years (Unix or Linux), which is dying. Their is also OpenCL, but have you question how many games run under OpenGL/CL? These games will run on any platforms. Did DirectX got killed by OpenGL? No. It is stronger than ever now. Point?

If the CUDA will die, then directcompute won't be far from it because they technically does the same thing. However, those who utilizes GPGPU probably don't know anything about DirectX as it is really for gaming purpose.