Can DX9 cards become run DX10 with a new driver?

[AlexV]

The current X3000 was never marketed as such, and no driver will be made to enable DX10 functionality for it, Intel will release another IGP for DX10.

 

[DavidC]

The current X3000 was never marketed as such, and no driver will be made to enable DX10 functionality for it, Intel will release another IGP for DX10.

Actually, it was. Perhaps its gonna be DX10 with GM965, the mobile version coming with Santa Rosa...

 

[Nick]

According to that paragraph then, the current GMA X3000 is not DX10 compatible: http://www.intel.com/cd/ids/developer/asmo-na/eng/recent/334680.htm?page=3

Why?? Cause the FP precision for the Pixel Shader is only 24 bit.

What are the precision requirements for DirectX 10 actually? NVIDIA's multi-function interpolator has an accuracy of 22 mantissa bits.

 

[Demirug]

According to that paragraph then, the current GMA X3000 is not DX10 compatible: http://www.intel.com/cd/ids/developer/asmo-na/eng/recent/334680.htm?page=3

These lists on this page need some proof reading. Shader model 3 already requires 32 bit precession.

 

[AlexV]

Can anybody point me to an official Intel marketing material where the X3000 is treated as a DX10 part?And Demi is correct, I think the X3000 is full FP32.

 

[JHoxley]

What are the precision requirements for DirectX 10 actually? NVIDIA's multi-function interpolator has an accuracy of 22 mantissa bits.

There are various storage formats (10bit, 11bit, 16bit, 32bit) but all computation is in a close subset of IEEE 754 32-bit single precision floating point specification. Although that requires a 23bit mantissa (+ 1 signed bit and 8 exponent bits) so I don't quite see how the referenced Nvidia document fits in...

The main deviations from IEEE 754 are the extreme cases - FP exceptions, NaN's, INF's etc... There also seems to be a few relaxations to improve performance (e.g. faster rounding versus accurate rounding) and aid conversions to/from integer formats.

hth
Jack

 

[Nick]

There are various storage formats (10bit, 11bit, 16bit, 32bit) but all computation is in a close subset of IEEE 754 32-bit single precision floating point specification. Although that requires a 23bit mantissa (+ 1 signed bit and 8 exponent bits) so I don't quite see how the referenced Nvidia document fits in...

I noticed that the DirectX 9.0c SDK documentation specifies 2^-21 precision for exp and log, and 2^-22 rcp and rsq. So I assume this hasn't changed for DirectX 10.

The main deviations from IEEE 754 are the extreme cases - FP exceptions, NaN's, INF's etc...

Obviously a GPU won't throw an exception, but the DirectX 9.0c SDK documentation specifies that rcp and rsq should return inf for 0.0 input. Properly handling NaN's also doesn't seem too hard.

There also seems to be a few relaxations to improve performance (e.g. faster rounding versus accurate rounding) and aid conversions to/from integer formats.

Yes, round to even and denormal rules are likely not followed since they require a significant amount of extra logic and latency.

It's a bit ironic though. The accuracy is overkill for graphics, but it's too low for truely scientific GPGPU applications. :|

 

[ET]

Actually with CTM/Cuda becoming available it might be possible to emulate the basic functionality fairly effectively.

While this point is meaningless for CUDA, since it's available only on cards which already support DX10, I agree for CTM. It has a features not exposed through DX9 (such as scatter writing), and could probably do a good job of offering DX10-like functionality.

The problem is, there's no such thing as "basic DX10 functionality", and emulating the full DX10 spec would be beyond what can be done on the X1x00. Besides, while an interim solution is nice to have, we'll have a good enough range of true DX10 solutions soon enough.