Game Dev's Video card

[tesseracter]

I've got the money to drop on a new video card, but need help deciding based on interesting factors. Ive got a Q9300, 8GB ram, 1920x1200, watercooled(gratuitous techness, but oh well) and run 85% Ubuntu, 15% windows.

I want to play with GPGPU, fluid simulation, and large amounts of particles. I want to have things I make run on other computers. I don't want to be caught with my pants down in terms of developing in the future. I want good API documentation.

only one PCI-E spot, so i need to make it count.

 

[RADEoN]

wait a few weeks, get the 4870X2

 

[kllrnohj]

Well, Nvidia currently has the better GPGPU implementation with CUDA, but ATI is publishing their cards specs so you'll see better Linux drivers (afaik both camps are more or less equal in current binary drivers).

I would have to recommend the GTX 280 as Nvidia is clearly trying to make CUDA mainstream and will likely succeed.

 

[cybereality]

I would go with the GTX 280. Nvidia is very good about providing free tools to developers, PhysX is free to use (even for commercial titles), and they always have the cutting edge. Obviously with an NV card you can run CUDA and there are actually a number of sample applications already available on Nvidia's developer page. I've worked with the PhysX API a bit and there is a decent amount of documentation. The SDK includes a large number of sample applications to get you started right off the bat. ATI is making some progress, but I still feel like they're some ways behind on the software development front.

People will argue price vs. performance, but at the end of the day Nvidia comes through with the horse-power. If you have the money you really can't go wrong with the GTX 280.

EDIT: I am only talking about Windows development, I don't really have any idea which is better in terms of Linux.

 

[flak-spammer]

If you want any kind of linux support you should go with Nvidia, ATI has a past history for very bad linux drivers (I havn't used one recently, but I do use linux and have had significant issues getting fglrx to work in the past). Not to mention Nvidia drivers have a much larger support across the linux side of things.

 

[Lost!]

OpenCL is *open*. It's not here yet, but for now Brook+ for the 4800s have improved quite a lot in performance.
CUDA is a really nice language to use ATM, considered the best GPGPU language because of consistency with drivers, but that is going both sides soon.

From what I read from Phoronix.com, nVidia's current drivers are stable, but they are slower in 2D. ATI's drivers for major builds are good, minor builds you might be out of luck (might be the same for nV, don't know). OpenSUSE being AMD supported has the easiest installation and handling of Radeons.

Something you need to know, though- There is NO de facto API for GPGPU now. It's a 3-way tug'o'war, and you shouldn't piggyback everything to a single language. The important part is the principles of coding for it, and not the language itself.

That said, for developing in CUDA I'd give the choice to the GTX 260. Just nice.

 

[Silus]

I've got the money to drop on a new video card, but need help deciding based on interesting factors. Ive got a Q9300, 8GB ram, 1920x1200, watercooled(gratuitous techness, but oh well) and run 85% Ubuntu, 15% windows.

I want to play with GPGPU, fluid simulation, and large amounts of particles. I want to have things I make run on other computers. I don't want to be caught with my pants down in terms of developing in the future. I want good API documentation.

only one PCI-E spot, so i need to make it count.

Right now, a CUDA ready GPU is the way to go. NVIDIA has tons of documentation in their site:

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

The GTX 280 is the most powerful GPU around, but when it was released, it was very expensive. With the recent price cuts however, it is even more appealing for about $500. If you have the money, go for it. If your budget is limited to something below $500, the GTX 260 can be had for about $350 or less nowadays.

 

[kllrnohj]

If you want any kind of linux support you should go with Nvidia, ATI has a past history for very bad linux drivers (I havn't used one recently, but I do use linux and have had significant issues getting fglrx to work in the past). Not to mention Nvidia drivers have a much larger support across the linux side of things.

Not anymore. ATI made a huge push on the Linux side of things. Their official drivers are now neck and neck with Nvidia's (nvidia has the edge in some games, but ATI has the better 2D), with a win going to ATI due to their open specs (meaning solid open source drivers).

 

[bacon]

Agreed the 4800's even ship with drivers. 2D is the suck with my 8800gt, and web pages tear so bad I can't use it for anything normal.

 

[kllrnohj]

Agreed the 4800's even ship with drivers. 2D is the suck with my 8800gt, and web pages tear so bad I can't use it for anything normal.

Enable vsync in the nvidia control panel, helped a ton with my 7950GT (although I was using Compiz, so it wasn't strictly 2d)

 

[2Fresh]

Agreed the 4800's even ship with drivers. 2D is the suck with my 8800gt, and web pages tear so bad I can't use it for anything normal.

Then you have no idea how to install a gfx card/drivers. The 8800gt in my work rig runs anything I play with nice settings....also has no trouble with "web pages"...

 

[bacon]

I have five linux boxes and four nvidia cards. The 8800gt is the only one that does this, and I am not alone. There are plenty with the same issues with G92 cards.

Not knowing how to install a driver would be an be possible if I wasn't a linux developer but somehow I think I can manage that. The 3d is fine, and works great but the 2d is jacked up.

No matter what module I use, if I compile it myself or use someones rpms/debs it just won't work.

The nvidia devs don't know what is causing it either, and you can do a search for yourself and see just how many people this is effecting.

The whole while gtk/qt/xorg and everything in between gets blamed. Everything since the 169.25 module has been bugged for me.

 

[tesseracter]

Thanks for all the responses.

looks like there isnt a clear winner, ATI linux drivers, nVidia CUDA(does cuda work in linux||openGL?) both 4870 and the 280 seem ok, same with 4870X2, is there anything that can make up my mind for me? a deal breaker? i dont really want to go to windows to program, so whatever behaves nicest in linux might take the cake.

 

[vengence]

If you want to play with multithread friendly problems, you'll want to play with CUDA. Try this link

http://www.nvidia.com/object/cuda_home.html#

If your worried about CUDA and Linux, I'd suggest heading over to the CUDA Linx forums:

http://forums.nvidia.com/index.php?showforum=68

 

[kllrnohj]

Thanks for all the responses.

looks like there isnt a clear winner, ATI linux drivers, nVidia CUDA(does cuda work in linux||openGL?) both 4870 and the 280 seem ok, same with 4870X2, is there anything that can make up my mind for me? a deal breaker? i dont really want to go to windows to program, so whatever behaves nicest in linux might take the cake.

CUDA works in *nix just fine so I would assume it uses OpenGL. You could always use like Brook+ for GPGPU work, and there is work on trying to run CUDA on ATI's hardware, so that is what I would lean towards. Either the 4870 or 4870X2

 

[vengence]

CUDA works in *nix just fine so I would assume it uses OpenGL. You could always use like Brook+ for GPGPU work, and there is work on trying to run CUDA on ATI's hardware, so that is what I would lean towards. Either the 4870 or 4870X2

If he wasnt' asking to be able to run GPGPU tasks then I would have recommened the 4870x2. But if he is interested in running multithreads then I would really suggest the 260 or 280. The Massive price drops of late have really changed where ther are on price/preformance. And I wouldn't want to buy somethign "hoping" CUDA would be implemented on it.

 

[kllrnohj]

If he wasnt' asking to be able to run GPGPU tasks then I would have recommened the 4870x2. But if he is interested in running multithreads then I would really suggest the 260 or 280. The Massive price drops of late have really changed where ther are on price/preformance. And I wouldn't want to buy somethign "hoping" CUDA would be implemented on it.

I wouldn't buy on "hope" either, but there are other GPGPU implementations, such as Brook+ (which has the advantage of running on older cards too). CUDA isn't the only one, just the most heavily advertised one

 

[vengence]

I wouldn't buy on "hope" either, but there are other GPGPU implementations, such as Brook+ (which has the advantage of running on older cards too). CUDA isn't the only one, just the most heavily advertised one

If the intent of playing with it is to put it on your resume, then using the most heavily advertised one is good! Remember, most managers are stupid!

 

[kllrnohj]

If the intent of playing with it is to put it on your resume, then using the most heavily advertised one is good! Remember, most managers are stupid!

Oh, of course, no doubt. But at the same time, all the solutions are very young, so I'm hesitant to try and make any long range predictions. That said, ATI has an official modified a version of Brook for their GPUs (Brook+), so its a tough call as to which would be preferred in the future: http://ati.amd.com/technology/streamcomputing/

 

[vengence]

Oh, of course, no doubt. But at the same time, all the solutions are very young, so I'm hesitant to try and make any long range predictions. That said, ATI has an official modified a version of Brook for their GPUs (Brook+), so its a tough call as to which would be preferred in the future: http://ati.amd.com/technology/streamcomputing/

If I were to guess, I'd guess CUDA. But thats just because it is being pushed so hard. But there isn't any way to know for sure.

 

[kllrnohj]

If I were to guess, I'd guess CUDA. But thats just because it is being pushed so hard. But there isn't any way to know for sure.

I think CUDA has potential, but it needs to get more widespread adoption (specifically it must run on ATI cards too) before it will really take off. There is also OpenCL in the works for open GPGPU spec.

 

[ZerazaX]

OpenCL is in the works and supposedly DX11 from Microsoft is going to have a Compute Shader component, so they might be introducing a DirectX API for GPGPU calculations, in which case any DX11 card should work on it so that might torpedo CUDA

 

[bassman]

CUDA works in *nix just fine so I would assume it uses OpenGL

Just a minor point, but CUDA doesn't "use" OpenGL. It doesn't "use" any API except the driver - the nvcc compiler generates code that runs on the stream processors directly after being interpreted by the runtime/driver.