NVIDIA claims top DX9 GPU marketshare spot.

[StealthHawk]

A waiver is effectively an area of non-compliance but MS are willing to overlook for certain reasons. I'm hearing that the 5200 has several waivers - we know that it only runs in FP16, so this is likely to be one of them.

Where did you hear that it runs only in fp16 mode?
(Obviously you don't "know" that it does since it does run fp32.)

Check DX9 caps.

With that said, I have heard that FP32 is once again enabled in 45.23 and 45.24 for the 5200, 5600, and 5800. Anyone here care to verify...or send me a gfFX card to test

 

[Clootie]

With that said, I have heard that FP32 is once again enabled in 45.23 and 45.24 for the 5200, 5600, and 5800. Anyone here care to verify...or send me a gfFX card to test

I can confirm what all FX cards support FP32 in PS_2_0 with new official 45.23 drivers, along with FP16 in partial precision (PP) mode.

 

[Dave Baumann]

Where did you hear that it runs only in fp16 mode?
(Obviously you don't "know" that it does since it does run fp32.)

Clearly it didn't with the 44.03's.

 

[digitalwanderer]

Where did you hear that it runs only in fp16 mode?
(Obviously you don't "know" that it does since it does run fp32.)

Clearly it didn't with the 44.03's.

I'm not disagreeing or questioning your statement here Dave, but how can you tell if it's doing fp32 or not?

 

[CorwinB]

I'm not disagreeing or questioning your statement here Dave, but how can you tell if it's doing fp32 or not?

I'd say that you can check if FP32 is available by running custom code generating visible artefacts in FP16 (Mandlebrot ?), and coded without specifying _PP...

 

[Tridam]

I'm not disagreeing or questioning your statement here Dave, but how can you tell if it's doing fp32 or not?

I'd say that you can check if FP32 is available by running custom code generating visible artefacts in FP16 (Mandlebrot ?), and coded without specifying _PP...

It's rather simple if you code your own shader. You can use a shader to mathematically check the epsilon (precision) and the range. The epsilon gives you the mantissa and the range gives you the exponent.

With det 44.03, FX5800, 5600 and 5200 were always doing FP16, with or without _pp. But this is not the case with det 45.23. Without _pp we have FP32.

 

[digitalwanderer]

I'm not disagreeing or questioning your statement here Dave, but how can you tell if it's doing fp32 or not?

I'd say that you can check if FP32 is available by running custom code generating visible artefacts in FP16 (Mandlebrot ?), and coded without specifying _PP...

It's rather simple if you code your own shader. You can use a shader to mathematically check the epsilon (precision) and the range. The epsilon gives you the mantissa and the range gives you the exponent.

With det 44.03, FX5800, 5600 and 5200 were always doing FP16, with or without _pp. But this is not the case with det 45.23. Without _pp we have FP32.

Oh, ok....I get ya now. You do some things that are totally above me head and beyond my current scope-o-understanding and find out that way.

How 'bout next time you just say "magic!", it's close enough for me little brain and it'll save you some typing.

 

[Reverend]

With that said, I have heard that FP32 is once again enabled in 45.23 and 45.24 for the 5200, 5600, and 5800. Anyone here care to verify...or send me a gfFX card to test

I can confirm what all FX cards support FP32 in PS_2_0 with new official 45.23 drivers, along with FP16 in partial precision (PP) mode.

I can verify that.

 

[Clootie]

It's rather simple if you code your own shader. You can use a shader to mathematically check the epsilon (precision) and the range. The epsilon gives you the mantissa and the range gives you the exponent.

Yes it's easy - Pixel Shader 2.0 precision

 

[Tridam]

It's rather simple if you code your own shader. You can use a shader to mathematically check the epsilon (precision) and the range. The epsilon gives you the mantissa and the range gives you the exponent.

Yes it's easy - Pixel Shader 2.0 precision

This software just check the registers and coordinates data type. It doesn't check the precision of the units !

IE : this software say that Radeon 9700/9800 use s23e8 registers. But every precision and range test will show you that Radeon do calculations in the s16e7 format. This software doesn't say that

[EDIT] : Sorry, it was an error of mine The software shows s16e7 for Radeons as expected.