Coder's Video Card?

[tesseracter]

Lets see what you people come up with for me.

I use ubuntu/linux almost exclusively, with a watercooled Q9300 and a quality 700W power supply. 24" LCD, 1920X1200 resolution.

I program games, and want to start messing with graphics card programming, fluid simulations, particle systems, all the stuff with using textures as datastructures. the basic reason for GPGPU in the first place.

I'm ready to buy a new video card, it will be water cooled so i dont card about the cooling solution that comes stock. I care about massive performance and futureproof programming on linux. does ATI or nVidia accel at either of these things?

 

[myren]

Right now its a hard call: theres too many factors. Sorry this is about to get terse. nVidia has classically had far better linux drivers but AMD started showing signs of life again recently and AMD is the only one with any hope at all of ever having viable open source drivers (just dont expect fully compliant fast 3d for another decade and a half from that). nVidia keeps crapping up their drivers though (now unplugging a monitor forcibly triggers xrandr events resizing your desktop) and imposes senseless artificial limitations on non Tesla/Quadro workstation cards. nVidia tries hard they just fuck up a lot and dont think about what the fuck they're doing: they try to give Linux users the same dogfood windows people get.

Sounds like you are interested in GPGPU work: nVidia has been backing their own CUDA and AMD currently has Brooks but has joined the OpenCL party that looks to be the standardized way of the future path and-- very importantly-- will be able to interop/share data with running OpenGL code. The OpenCL spec isnt released yet but I'm expecting far faster turn around from AMD. Both have commited to the much improved OpenGL3 spec which is excellent but historically ATI has been very weak in commitments to OpenGL.

nVidias SDK is far more demonstrative and well composed. Their website is breathtaking: they clearly want to show you how to abuse their systems. I dont know the state of nVidias tooling in Linux but in Windows its light years ahead.

The 48xx series has far far far far better double precision float performance. AMD has a tesselator unit that complements the geometry shaders very well and can provide immense performance boosts.

 

[myren]

For the record I'm buying a 4850 to put in my colo server. I'll take my chances with the crap sdk and slower Linux driver releases.

 

[topdog5050]

Linux get Windows go all the way.

 

[topdog5050]

Great post myren.

 

[Dan_D]

NVIDIA has a history of better Linux drivers so that's something to look at. Programers like Scali2 are generally impressed with CUDA on NVIDIA graphics cards as well. Apparently they are monsters when it comes to general processing type applications.

You'd want to talk to someone like that about it. I'm just relaying my impressions and memories of certain threads.

 

[tesseracter]

thanks myren.

so nVidia looks like its holding top place, but ATI is accelerating towards linux faster. so the thing i wonder about is: if i get a 4870 or 4870x2, is firmware enough to be compatible with openGL3, openCL, and such, or are those hardware things? if no cards have those things, i'll probably go with a 260 or something, but i dont buy video cards that often(4-5 years?)

 

[vengence]

A side not for you tess. CUDA does not require SLI, which means you can run 2- 260s on a non-SLI mobo down the road for CUDA.

From the stand point of user support and programming sensibility, CUDA has made it's programming VERY easy, yet still very powerful.

If I were you, I'd take a look around Nvidia's cuda site. You might even go ahead and download the SDK, just to read through the documentation.

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

 

[kllrnohj]

For the record I'm buying a 4850 to put in my colo server. I'll take my chances with the crap sdk and slower Linux driver releases.

ATI's Linux release schedule is now the same as the Windows one. Cat 8.8s, for example, came out on Linux the same day they did on Windows. They also had same-day support for the 48xx series (although the download page didn't list the 48xx cards)

thanks myren.

so nVidia looks like its holding top place, but ATI is accelerating towards linux faster. so the thing i wonder about is: if i get a 4870 or 4870x2, is firmware enough to be compatible with openGL3, openCL, and such, or are those hardware things? if no cards have those things, i'll probably go with a 260 or something, but i dont buy video cards that often(4-5 years?)

No, ATI currently holds the top spot in Linux drivers. They successfully passed up nVidia a couple of months ago in terms of features and support (performance is a bit of a draw. Nvidia leads in gaming performance, but ATI has the edge in 2d performance). Nvidia had previously been dominating the linux driver area, but after the AMD acquisition of ATI that started to turn around. Now ATI has excellent Linux drivers, is opening up their specs for open source drivers, has the same release schedule as the Windows drivers, even has OverDrive support in the drivers now. The ATI drivers also build packages for the distro's package manager, a major plus.

You really can't go wrong with either card from a driver standpoint. In terms of GPGPU, CUDA seems to be better supported than AMD's Stream SDK at the moment...

 

[myren]

tesseract, both pieces will be OpenGL3 compliant (with commitments from both manufacturers) and I'd expect all the modern hardware to support OpenCL.

Especially given the driver support kllrnohj mentioned, Cuda would be the big nvidia draw right now (and its huge) but I think this will equalize. As I've stated, right now Cuda is far superior, simply because nVidia has done such an excellent job of supporting it, but I strongly believe the future is all about the upcoming open standard: OpenCL. Both companies will definitely be able to run OpenCL and if nVidia' SDK ends up migrating to OpenCL the same demos and tutorials will be applicable to both. But if you want to develop GPGPU NOW Cuda is really away and above.