What GPGPU applications are available to ATI users?

[GaiaHunter]

But you're right, I should dump the Radeon and get a Fermi. I've thought about it. I just didn't think Fermi was a good enough product. But every month that passes without AMD adding OpenCL to their drivers, Fermi gets more attractive.

But isn't NVIDIA exactly in the same position as AMD?

Or are these R195 drivers the same as consumers drivers?

http://www.nvidia.com/object/cuda_opencl_new.html

http://developer.amd.com/gpu/atistreamsdk/pages/default.aspx

 

[Scali]

But isn't NVIDIA exactly in the same position as AMD?

No.

Or are these R195 drivers the same as consumers drivers?

As I said, consumer drivers since November:
http://www.nvidia.com/object/win7_winvista_32bit_195.62_whql.html
"Adds support for OpenCL 1.0 (Open Computing Language) for all GeForce 8-series and later GPUs."
Check all release notes if you like, they all mention OpenCL. Or install them and try to run OpenCL code (eg the samples from GPU Caps Viewer). You'll see that they work out-of-the-box.

Note also that this is for all GeForce 8-series!
AMD only supports the 4000 and 5000-series, not the 2000 and 3000-series, which date from the same era as the 8-series.

 

[grimpr]

I'm quite sure that they will, actually. Intel supports the OpenCL standard: http://software.intel.com/en-us/blogs/2009/01/21/parallel-programming-talk-opencl-with-tim-mattson/
As for DirectCompute, technically their current IGPs should already be able to support it (albeit at CS4.0 level only). They haven't enabled it in the drivers yet, but I'd be surprised if they don't enable it on Sandy Bridge.

You seriously believe this? Sandy Bridges IGP supporting OpenCL/DirectCompute/OpenGL 4.0/DX11 etc...

 

[Scali]

You seriously believe this? Sandy Bridges IGP supporting OpenCL/DirectCompute/OpenGL 4.0/DX11 etc...

Where did OpenGL 4.0 and DX11 come from? I didn't say anything about that.
I only said OpenCL and DirectCompute... which as I said, I don't see why not... CS4.0 should be doable. And that's about equal to OpenCL 1.0. Even the nVidia G80 supports those two technologies.

 

[GaiaHunter]

So for AMD GPUs I need the 10.4 Cat drivers to have OpenCL.

Do I need the opencl driver which is (72 MB for 32-bits OS and 120 MB for 64-bits OS) as well?

And if I need both the big problem is a 120 MB max download?

 

[Scali]

Do I need the opencl driver which is (72 MB for 32-bits OS and 120 MB for 64-bits OS) as well?

There is no driver, only a complete SDK.

And if I need both the big problem is a 120 MB max download?

Not really...
The problem is that an end-user doesn't know what an SDK is, let alone where to get it... and why would they have to install a complete SDK, which they don't know what to do with anyway? They just need an OpenCL runtime, which could just be bundled with the driver, like nVidia does. That way the end-user doesn't need to do anything other than a driver update. Something that most of them should already be familiar with, either manually, or via Windows Update.

 

[GaiaHunter]

Not really...
The problem is that an end-user doesn't know what an SDK is, let alone where to get it... and why would they have to install a complete SDK, which they don't know what to do with anyway? They just need an OpenCL runtime, which could just be bundled with the driver, like nVidia does. That way the end-user doesn't need to do anything other than a driver update. Something that most of them should already be familiar with, either manually, or via Windows Update.

They click on the .exe as they do with the drivers? Or even better the application that they are installing does that for them?

It is a bit similar as when games installed DX or when games install the physX drivers, as in for example Dragon Age.

People don't care about installing the SDK because there is basically no use for OpenCL yet.

What application are you developing for end users?

 

[waffleironhead]

There is no driver, only a complete SDK.



Not really...
The problem is that an end-user doesn't know what an SDK is, let alone where to get it... and why would they have to install a complete SDK, which they don't know what to do with anyway? They just need an OpenCL runtime, which could just be bundled with the driver, like nVidia does. That way the end-user doesn't need to do anything other than a driver update. Something that most of them should already be familiar with, either manually, or via Windows Update.

I prefer the graphics driver to be as small as possible. The SDK or runtimes should be included with the application imo.

 

[Voo]

I prefer the graphics driver to be as small as possible. The SDK or runtimes should be included with the application imo.

You want a dozens of applications to include the OpenCl runtime - probably dated and getting you into troubles which one to use - instead of one, a few MB larger download? I don't even now if applications are allowed to bundle the runtime legally..

Sounds like a bad idea compared to just bundling it with the drivers, I mean who cares today about 20mb more or less?

 

[Scali]

You want a dozens of applications to include the OpenCl runtime - probably dated and getting you into troubles which one to use - instead of one, a few MB larger download? I don't even now if applications are allowed to bundle the runtime legally..

Sounds like a bad idea compared to just bundling it with the drivers, I mean who cares today about 20mb more or less?

Worse than 'the runtime'... you need to include the runtime for all common OpenCL devices... that includes AMD, nVidia, Intel, VIA/S3... possibly others?
Nope, never going to work, they just need to bundle it with the graphics driver, where it belongs.

 

[GaiaHunter]

Bundle it with the driver or just have it there for those that want to download it... which they already have it.

Might be more practical to have it bundle with the GPU drivers but I hardly call it something that deters anything. A single informative html pointing to the right place would be enough.

 

[Scali]

Bundle it with the driver or just have it there for those that want to download it... which they already have it.

Might be more practical to have it bundle with the GPU drivers but I hardly call it something that deters anything. A single informative html pointing to the right place would be enough.

It creates a chicken-and-egg problem that is unneccessary. There's no reason why AMD shouldn't bundle it with their drivers.
nVidia does it, S3 does it.
The only reason why AMD doesn't is simple: they'll get creamed by nVidia as soon as people can do apples-to-apples application benchmarks in OpenCL.

 

[GaiaHunter]

It creates a chicken-and-egg problem that is unneccessary. There's no reason why AMD shouldn't bundle it with their drivers.
nVidia does it, S3 does it.
The only reason why AMD doesn't is simple: they'll get creamed by nVidia as soon as people can do apples-to-apples application benchmarks in OpenCL.

But everyone can run benches - download 10.4 drivers and the SDK - took me 5 minutes.

 

[Scali]

But everyone can run benches - download 10.4 drivers and the SDK - took me 5 minutes.

In theory, yes... but not many relevant benches exist yet. Partly because AMD was able to control the situation so far. And partly because AMD shot themselves in the foot and drove developers back to Cuda.

 

[GaiaHunter]

In theory, yes... but not many relevant benches exist yet. Partly because AMD was able to control the situation so far. And partly because AMD shot themselves in the foot and drove developers back to Cuda.

Except you - you want to develop for OpenCL and not CUDA.

 

[Scali]

Except you - you want to develop for OpenCL and not CUDA.

If at all possible, I prefer to support the technology that supports the widest range of hardware and thus the largest demographic, yes.
But currently, OpenCL is pretty much only nVidia and S3, and the S3 marketshare is negligible... so the difference in marketshare with Cuda is marginal, while C for Cuda is easier to use and generally a more powerful tool than OpenCL.
If it wasn't for the fact that I don't have an nVidia card in my machine right now, I might have gone back to Cuda already... who knows.

 

[GaiaHunter]

If at all possible, I prefer to support the technology that supports the widest range of hardware and thus the largest demographic, yes.
But currently, OpenCL is pretty much only nVidia and S3, and the S3 marketshare is negligible... so the difference in marketshare with Cuda is marginal, while C for Cuda is easier to use and generally a more powerful tool than OpenCL.

How the hell is OpenCL only NVIDIA if you can program it to run on AMD hardware and I can use it if you do so?

And still, CUDA has those exact same consumer applications + adobe and maybe something else.

GPGPU is still at infant stage - one can do a few steps while the other can only move on 4.

 

[Voo]

If it wasn't for the fact that I don't have an nVidia card in my machine right now, I might have gone back to Cuda already... who knows.

Well you can use their emulator (though afaik that's already deprecated), which works.. well it works and if you want to debug some code that's the most sensible approach (though coding with GPUs really reminds you to look at sourcecode and not just start up some debugger - oh and all those IFDEFS for the emulator..), but yeah without a cheap GPU it's probably problematic.

 

[GodisanAtheist]

I think what ScholzPDX is looking for, is for AMD's GPGPU to be at a point where AMD aficionados don't feel as though they have to defend it.

 

[Scali]

I think what ScholzPDX is looking for, is for AMD's GPGPU to be at a point where AMD aficionados don't feel as though they have to defend it.

Very good point.

 

[Scali]

How the hell is OpenCL only NVIDIA if you can program it to run on AMD hardware and I can use it if you do so?

You can run OpenCL perhaps, I can run OpenCL, but the average end-user can't.
You think the average end-user for PhotoShop/Premiere has any idea about GPGPU acceleration or what Cuda really means? Let alone that they would know where to find an SDK and install it?
I doubt it, these are artists, not computer-savvy people (most of the time they use Macs for that reason).
All they have to know is that Adobe recommends nVidia Quadro cards for best performance, so they buy one of those, and probably have it installed for them by the IT department.

Let me explain in terms you may understand (but probably will deny anyway):
AMD advertises with OpenCL support. So people who buy an AMD card, will think that it supports OpenCL. Then they install my application... and hey, it doesn't work!
Who do you think they're going to call for support, me or AMD?
I'd be getting a lot of support calls saying "Hey your software is crap, it doesn't work", and then I'd have to explain that they need to download and install the SDK etc.
Yes, I could put it in the manual, in a readme, or even a popup window on install... but do you think people actually read those? Let me answer that for you: They don't.
So it'd be a big waste of time and money on my behalf. Thank you AMD.

GPGPU is still at infant stage - one can do a few steps while the other can only move on 4.

No, AMD's GPGPU is still at its infant stage. nVidia on the other hand is doing fullblown C++ by now. With nVidia there's barely a difference between CPU and GPU anymore, language-wise. nVidia even offers Visual Studio plugins for debugging and performance analysis.

 

[GaiaHunter]

No, AMD's GPGPU is still at its infant stage. nVidia on the other hand is doing fullblown C++ by now. With nVidia there's barely a difference between CPU and GPU anymore, language-wise. nVidia even offers Visual Studio plugins for debugging and performance analysis.

Again, fantastic.

Now convince me to buy a NVIDIA card to run all my CUDA applications.

Tell me about this application that is based on CUDA that I can't simply live without once I experience it, that it will change the way I use my PC.

The fact is, you keep talking how amazing NVIDIA GPGPU is and how shitty AMD is (no references to Intel). And it might just be as you say.

But where are those applications that will make my PC so different, that will enable me to do stuff I don't do today?

I bet somehow is AMD fault those applications aren't here.

Just like physX, it is great technology is going to be the future for the last 3 years and I still haven't seen anything that makes me say "There is only 1 option for my GPU and that is NVIDIA cause there is no way I'm not having physX".

 

[Scali]

Now convince me to buy a NVIDIA card to run all my CUDA applications.

That isn't the point.
The technology is mature, but that doesn't mean that regular consumers necessarily benefit from it at this point.
Cuda is mainly in use in the HPC/scientific world, where it is a very cost-effective alternative to conventional supercomputers.

Just like physX, it is great technology is going to be the future for the last 3 years and I still haven't seen anything that makes me say "There is only 1 option for my GPU and that is NVIDIA cause there is no way I'm not having physX".

This argument is even poorer than the "I don't see any GPGPU applications that I would like to use".
Games like Batman AA and Mafia II prove that PhysX can add some nice eyecandy and improve realism quite significantly.
I'm getting tired of AMD fanboys who are still in denial about that.
And those same AMD fanboys were crapping on nVidia for not having DX11, which has done far less for eyecandy and realism than the few PhysX games that have come out so far.
And now with nVidia's Fermi, you can have your cake and eat it too: DX11 and PhysX (and Cuda, and DirectCompute, and OpenCL etc).
And what is AMD doing on the physics front in the meantime? Nothing. There's no sign that AMD will ever offer an alternative to PhysX. I guess that's why you're in denial about PhysX in the first place, right?
Any objective observer would see value in the added realism, and would see the potential of the technology once it becomes commonplace, and CPU-based workarounds are no longer necessary. The rest are just fools.

Accelerated physics aren't an just an nVidia thing. It's far bigger than that. It's the next big step in gaming realism. The first big step since the first 3d accelerators arrived (I've been saying that ever since the first Ageia PPUs arrived, long before there was any sign of nVidia adopting it for their GPUs).
Don't be an idiot and make an nVidia vs AMD thing out of this. You're missing the bigger picture.

 

[Voo]

The fact is, you keep talking how amazing NVIDIA GPGPU is and how shitty AMD is (no references to Intel). And it might just be as you say.

The only thing he said is that it's much easier to develop for CUDA than OpenCL, nothing else. And if you disagree I'd love to see what programs you've written with those two..

 

[GaiaHunter]

That isn't the point.
The technology is mature, but that doesn't mean that regular consumers necessarily benefit from it at this point.
Cuda is mainly in use in the HPC/scientific world, where it is a very cost-effective alternative to conventional supercomputers.

That is why the regular costumers, like me, don't give a damn and you shouldn't be offended by it.

This argument is even poorer than the "I don't see any GPGPU applications that I would like to use".
Games like Batman AA and Mafia II prove that PhysX can add some nice eyecandy and improve realism quite significantly.

And other games have as good or better eye candy and look as real or more.

I'm getting tired of AMD fanboys who are still in denial about that.
And those same AMD fanboys were crapping on nVidia for not having DX11, which has done far less for eyecandy and realism than the few PhysX games that have come out so far.
And now with nVidia's Fermi, you can have your cake and eat it too: DX11 and PhysX (and Cuda, and DirectCompute, and OpenCL etc).

And more money spent, and more power consumption, and more heat and more noise.

And what is AMD doing on the physics front in the meantime? Nothing. There's no sign that AMD will ever offer an alternative to PhysX. I guess that's why you're in denial about PhysX in the first place, right?

In the 3 years that I've been in denial you have Batman, you have Mirrors Edge, you have 2-3 other games that are passable and you will have Mafia II.

Any objective observer would see value in the added realism, and would see the potential of the technology once it becomes commonplace, and CPU-based workarounds are no longer necessary. The rest are just fools.

Sincerely added realism for games have its pros and cons. Of course that extra added realism also eats a good chunk of the frame rate and/or requires an additional GPU.

And again, physX, just like CUDA, hasn't become common place.

Accelerated physics aren't an just an nVidia thing. It's far bigger than that. It's the next big step in gaming realism. The first big step since the first 3d accelerators arrived (I've been saying that ever since the first Ageia PPUs arrived, long before there was any sign of nVidia adopting it for their GPUs).

It is the next big step. See, is the next big step. Future. Not happened yet. We agree.

Don't be an idiot and make an nVidia vs AMD thing out of this. You're missing the bigger picture.

My bigger picture is quite simple:

Instead of buying an Ageia PPU or a Fermi, I've bought something that is cheaper and uses runs today features. In the future I'll buy something to run the future features.

I understand you mate, all this great technology, very mature, that is the future and insist on staying in the future and not materializing.

AMD is really lucky, isn't it?

By the way, one of the Ageia guys that was working at NVIDIA seems to think the future is Fusion. Ironic, isn't it?