Will WGF 2.0 video cards be usable in XP/2000?

[KimB]

No, that's not what he said at all. All that they need to do is increase the size of the register file.

 

[Bob]

No, that's not what he said at all. All that they need to do is increase the size of the register file.

Increasing the register file costs area on-chip. A lot of it. Wouldn't you rather have those millions of transistors spent on better compression schemes, new features, or more MADs?


It's not because something is in some arbitrary API that it's automatically a good idea.

 

[KimB]

Right. All I'm saying is that there's not necessarily any need to worry about FP32 performance on a DX10 NV part (assuming partial precision is gone...is it really?). Obviously there is some optimal balance between register file size and number of execution units.

 

[JHoxley]

assuming partial precision is gone...is it really?

The current FX10 compiler still has the PP switch, but I'm pretty sure thats just for downlevel targets and that it's not valid for SM4.

From what I've seen there is no reason why NV/ATI couldn't have FP16 or FP24 somewhere as long as it was bypassed/skipped whenever any D3D10 code came down the pipe. If it was active during D3D10 usage then it would almost certainly make it a non-compliant D3D10 part.

Whilst D3D10 might be the primary reason for this next-gen hardware, it's not the only API that will make use of it (OpenGL and D3D9 for example) - so if those API's can make enough of a case for PP floats then maybe they'll dedicate some hardware for it. As I mentioned earlier - I doubt this would be the case though.

Cheers,
Jack

 

[Demirug]

The current FX10 compiler still has the PP switch, but I'm pretty sure thats just for downlevel targets and that it's not valid for SM4.

I am just run a test last night. I have used the compiler to generate bytecode and then feed the dissembler with it. Even if I am used the half datatype the shader bytecode does only contains floats.

 

[JHoxley]

I am just run a test last night. I have used the compiler to generate bytecode and then feed the dissembler with it. Even if I am used the half datatype the shader bytecode does only contains floats.

Even for Shader Models 1, 2 and 3? I've not tried any downlevel compiling in fxc10 - only SM4 stuff for me

My interpretation of available info was that, at least for now, fxc10 will correctly compile previous shader models in the same way the current fxc does.

Cheers,
Jack

 

[Demirug]

Even for Shader Models 1, 2 and 3? I've not tried any downlevel compiling in fxc10 - only SM4 stuff for me

My interpretation of available info was that, at least for now, fxc10 will correctly compile previous shader models in the same way the current fxc does.

Cheers,
Jack

I am not used fxc10. I have done it with D3D10CompileShader ( from C# )