Doomtrooper
Veteran
Alright Pascal 
Help the little guy
Help the little guy
Where do you stand with the FX
- Thread starter Typedef Enum
- Start date
"non-standard" as in IEEE.
R300 = 4channels x 24bit
NV30 = 4channels x 32bit
128-bit color = 4 32-bit IEEE floating-point values
I believe I wrote the following here in the forums :
You need genuine 32-bit floating point in order to perform serious per-pixel lighting math in the pixel shader, where you might want to treat per-pixel components as positions in space, 4x4 homogeneous matrices, etc., and transform them according to some physically-based lighting model. Some proprietary 24-bit floating point format would be fine for representing colors and normal vectors, but it would totally rule out using pixel shaders for any kind of positional computations.
The R300, even though it supports input and output of 32bit IEEE SPFP per component textures to the shader, it converts the external format to an internal 24bit format in the pixel shader, and convert it back before output. Given my examples above, problems may occur (when apps that expect "32-bit floating point" to mean "32-bit floating point").
Can anyone tell me if DX9's definition of 32bit floating point pixel spec isn't "a floating point format with an unspecified number of bits that might be more than 16 but could be less than 32"?
R300 = 4channels x 24bit
NV30 = 4channels x 32bit
128-bit color = 4 32-bit IEEE floating-point values
I believe I wrote the following here in the forums :
You need genuine 32-bit floating point in order to perform serious per-pixel lighting math in the pixel shader, where you might want to treat per-pixel components as positions in space, 4x4 homogeneous matrices, etc., and transform them according to some physically-based lighting model. Some proprietary 24-bit floating point format would be fine for representing colors and normal vectors, but it would totally rule out using pixel shaders for any kind of positional computations.
The R300, even though it supports input and output of 32bit IEEE SPFP per component textures to the shader, it converts the external format to an internal 24bit format in the pixel shader, and convert it back before output. Given my examples above, problems may occur (when apps that expect "32-bit floating point" to mean "32-bit floating point").
Can anyone tell me if DX9's definition of 32bit floating point pixel spec isn't "a floating point format with an unspecified number of bits that might be more than 16 but could be less than 32"?
I'm confused.
"Some proprietary 24-bit floating point format would be fine for representing colors and normal vectors, but it would totally rule out using pixel shaders for any kind of positional computations. "
I thought the precision of processing was 24-bit, but it could be represented to the shaders as 32-bit? I thought the only issue here was a greater error accumulation during processing?
"Some proprietary 24-bit floating point format would be fine for representing colors and normal vectors, but it would totally rule out using pixel shaders for any kind of positional computations. "
I thought the precision of processing was 24-bit, but it could be represented to the shaders as 32-bit? I thought the only issue here was a greater error accumulation during processing?
Tahir said:
When any benchmark app is produced in co-operation with a specific IHV is it likely that an optimised code path for that IHV was developed also?
I can acccept it is showing show fillrate advantages. What other reliable fill-rate intensive benchs are there? Rev suggested it was PS advantages. I suggest it is code path as well.
Apparently gainward will not use the FXFlow and will be quite silencious...
In french sorry
In french sorry
http://www.hardware.fr/news/lire/30-01-2003/#5524
Joe DeFuria
Legend
As I said in another post, I expect some vendors to provide "alternative" cooling methods. At the same time, I expect such products to be available at a later date, and also more expensive than the "DustBuster" FX models.
Can't exactly tell from the Gainward mail I got (poor format for the web mail reader), but it seems their Ultra model will be priced at £650 (inc VAT)!!
I'll update when I can read it properly.
[Edit] Its 650 Euro's.
I'll update when I can read it properly.
[Edit] Its 650 Euro's.
Joe DeFuria
Legend
If nothing else, this takes the prize for the worst product name in history:
Gainward FX PowerPack! Model Ultra/1000 Plus "Golden Sample" TM
Gainward FX PowerPack! Model Ultra/1000 Plus "Golden Sample" TM
Similar threads
