The official GT300/Fermi Thread

[flippin_waffles]

Well why stop at gt300? lol Estimates put gt300 3-6 months away. i'll suggest that in another 3-6 months after that, their will be something well worth your time waiting for. so really, you might as well wait another 6-12 months, unless what you really want to say is "i'd rather own an nv card". if so, just grow some balls and say it.

.

[grimREEFER]

wait a minute, if nvidia can do something complex like tesselation via cuda, what the hell is gonna stop gt300 from supporting every future api via cuda?

[Mad1723]

what the hell is gonna stop gt300 from supporting every future api via cuda?

The fact that shaders change considerably each time a new API comes out, that there are new requirements for precision and calculation capabilities, new compression algorithms, bigger textures... lots of stuff changes and it wouldn't be efficient to keep it as is, the performance hit of having programmable shaders doing specialized shaders stuff would probably be pretty high.

Then again, it could be possible, I'm not a specialist in any, I could be wrong.

[Farinorco]

wait a minute, if nvidia can do something complex like tesselation via cuda, what the hell is gonna stop gt300 from supporting every future api via cuda?

Not via CUDA but via shaders. I haven't took a more or less in-depth view of DX11, but as far as I have seen, I think there are 2 new types of shaders that allow to program the tesselation, Domain and Hull Shaders (in addition to the previous Pixel, Vertex and Geometry Shaders). That's not something especif to NVIDIA, but the way DX11 is defined.

Anyway, via CUDA (i.e., via GPGPU, be it CUDA, OpenCL, ATI Stream or what you want) you can effectively program an entire rendering process from scratch, using whatever aproach you want, and modeling the rendering pipeline to your convenience.

The downside? You would be programming everything to be run on general compute processors. The reason why even today we use Direct3D/OpenGL with their mostly fixed pipeline it's because the hw implements part of the tasks with hw especif to them (there are units to map the 2D textures to vertices of the 3D meshes, to apply filters, to proyect the 3D data to a 2D bitmap created by the fustrum of the camera, and so on). All this work is (logically) done much faster with hw with the especific mission to do it (TMUs and ROPs basically).

But yeah, I think the future of the 3D graphics will be on completely programmable pipelines, and the especific hw units to do particular tasks will disappear. When there's enough power to allow it, of course. That's the general direction with computers. The more power, the more we tend to flexibility.