R350 supports 128-bit full floating point precision????

[engall]

Maximum 3D
SMARTSHADER™ 2.1 supports the latest Microsoft® DirectX® 9.0 feature set, enabling cinematic quality lighting and shadow effects.
SMOOTHVISIONâ„¢ 2.1 delivers the sharpest, clearest textures without sacrificing frame rates.
128-bit full floating point precision enables billions of color variations to render the same lighting and effects as Hollywood studios.
Featuring CATALYSTâ„¢, the industry's most stable 3D acceleration software.
http://www.ati.com/products/radeon9800/radeon9800pro/features.html

 

[horvendile]

I believe that would be external precision.

 

[engall]

I believe that would be external precision.

Anything different?

 

[Simon F]

128-bit full floating point precision

I would guess that a comma is missing there...
"128 bit, full floating point" i.e. four colour channels each 32bit.

 

[Ailuros]

I believe that would be external precision.

Anything different?

Between read and output I'd say yes.

 

[KimB]

Maximum 3D
SMARTSHADER™ 2.1 supports the latest Microsoft® DirectX® 9.0 feature set, enabling cinematic quality lighting and shadow effects.
SMOOTHVISIONâ„¢ 2.1 delivers the sharpest, clearest textures without sacrificing frame rates.
128-bit full floating point precision enables billions of color variations to render the same lighting and effects as Hollywood studios.
Featuring CATALYSTâ„¢, the industry's most stable 3D acceleration software.
http://www.ati.com/products/radeon9800/radeon9800pro/features.html

The R3xx architecture reads and writes at 32-bit FP per channel (for 128 bits total). However, everything is stored on-chip at 24-bit FP. The use of 32-bit FP for storage is purely for memory performance reasons (it's easier to deal with power-of-two size blocks).

In other words, it is useless to think of the Radeon 9500+ cards as supporting anything higher than 24-bit FP. They may read and write more bits, but that means nothing. Its floating-point processing precision is 24-bit.

 

[Pete]

Well, the R3x0 pixel shaders are 24-bit, but the vertex shaders are 32-bit ... right?

 

[OpenGL guy]

Well, the R3x0 pixel shaders are 24-bit, but the vertex shaders are 32-bit ... right?

Yes. 32-bit floating point is required for VS 2.0.