FP16 and market support

[cthellis42]

At best, the people arguing single precision should atleast argue that FP32 needs to be supported as the minimum. If ATI indeed did not "finalize" their HW until MS said "FP24 is minimum", and if ATI indeed could have supported more than FP24 "easily", why didn't they move to FP32 in the beginning and lobby Microsoft to define FP32 as the minimum. At the time, NV already had FP32 capable HW (although with pathetic performance), and ATI could have delivered FP32 "easily", so there could have been consensus on FP32. Therefore if MS had endorsed FP32 as the minimum, we wouldn't have to wait another generation for the standard to be bumped up, since both vendors could have had FP32 HW ready, and ATI still would have come out looking golden, because presumably, their FP32 implementation would have "wiped the floor" with NVidia.

A) How would they know exactly what NV30 would look like?
B) Why increase silicon and decrease yields for something that may not get used fully--or even well--for a generation?
C) Wouldn't we still be in the SAME place as we are right now, except with developers programming their long shaders to FP32? R3xx would still be running them straight, and NV3x would still be mix-moding and _pp'ing.

Plus, there are more players than ATi and nVidia who probably didn't want the added complexity and didn't see the need for it straight off. And who knows how many developers felt they needed to go to that precision as minimum while they're just starting to roll out those kinds of advanced shaders? As well, we already saw card generations split into DX8 and DX8.1, and PixelShader go through many sub-steps, so why would folks fret overly at similar stepping in DX9? (Especially since by all accounts we're going to ride on it for a while.)

 

[Doomtrooper]

Amazing that 1.5 years ago, FX12 was fine, some of the most popular games of all time never even came close to requiring Pixel Shaders period, and in that short time people actually think that the complexity of developing a graphic card to support FP32 with speed would be a reality.
The move for FP24 as a minimum was a intelligent one, a stepping stone to something better, many titles like Doom 3 which I see used alot in the 'who is better' arguement won't even be using HDR, but rather FX16 precision which any decent DX8 accelerator supports.
We are talking 'gaming' cards here...it is always nice to speculate..yet I have not seen anyone provide enough evidence that FP24 would not provide the proper range in a majority of popular titles. Sure some extreme rare cases that are laughable at best.
If a developer can't make a game look good with FP24, when they have been doing it with FX12 for all these years...there is a problem.

 

[KimB]

Amazing that 1.5 years ago, FX12 was fine, some of the most popular games of all time never even came close to requiring Pixel Shaders period, and in that short time people actually think that the complexity of developing a graphic card to support FP32 with speed would be a reality.

1.5 years ago, FX12 was not fine (not for all operations). Texture ops on the NV2x were still done at FP32.

 

[Dave Baumann]

The FX9 combiners were still somewhat the limitation, as shown by the emboss comparisons between R200 and NV2x.

 

[YeuEmMaiMai]

well now that we have some preliminary stalker benchmarks we can see that once again ATi is clearly dominating nVidia here

1024*768 pure speed danger demo

Radeon 9800TXT 249.7
GForce 5950Ultra 227.5

so already it looks like in DX9 game that has been extensively optimized for Nvidia (according to the developers), they are down by 22.2fps (8.93%) despite having a 38Mhz clockspeed advanage....

enable 4x ansio and 4x FSAA and it gets even worse

Radeon 9800XT 183.1
Gforce 5950Ultra 138.3

now we are talking about 44.8 frames or 24.47%


Other demos are similar some are closer and some show even wider gaps between the ATi and Nvidia solutions

ALso JC stated that when you run ARB2 path R300 appears to be twice as fast as NV30 so I fail to see where you say what I stated above is not true........

 

[Daliden]

ALso JC stated that when you run ARB2 path R300 appears to be twice as fast as NV30 so I fail to see where you say what I stated above is not true........

And all this time I've thought that your gist was that using FP16 would result in a loss of image quality, which claim you still have not supported with quotes. I would assume that there isn't any in the S.T.A.L.K.E.R. game either (unless it's for Ati, if those rumors about FX-only shaders are true).

Yes, FX is slower than Radeon, but using FP16 keeps it at playable framerates and does not seem to result in loss of IQ. Notice that filtering tricks are a separate issue altogether and don't really have anything to do with shader precision.

 

[K.I.L.E.R]

YeuEmMaiMai, I don't think anyone is arguing that the FX series is better than the R3xx series.
It's great to see the R350 with better IQ tearing apart the NV35 but I think the main point is that you have to prove where FP16 results in less IQ under the current games than using precision > FP16.

 

[StealthHawk]

How are they going "backwards" with regards to AA? The gfFX has the same AA as the gf3 or gf4. I will agree that forcing brilinear is a step backwards quality-wise.

I thought the GFFX AA with it's weird "loopback blurring" was worse than on the previous generation, which IIRC, used supersampling. If you're suggesting that the previous Nvidia generation of AA is just as bad as on the GFFX, then I stand corrected. It's still not something to be proud of, especually when compared to their primary competition.

The previous generation(NV2x) used multisampling, although 4x supersampling is a hidden option that was exposed in the drivers sometime in 2003. NV1x used supersampling exclusively.

 

[KimB]

How are they going "backwards" with regards to AA? The gfFX has the same AA as the gf3 or gf4. I will agree that forcing brilinear is a step backwards quality-wise.

I thought the GFFX AA with it's weird "loopback blurring" was worse than on the previous generation, which IIRC, used supersampling. If you're suggesting that the previous Nvidia generation of AA is just as bad as on the GFFX, then I stand corrected. It's still not something to be proud of, especually when compared to their primary competition.

Only the Quincunx and 4x9 modes added any blurring. The normal AA modes never added blurring.

Nearly all FSAA modes have been unchanged since the NV20 from nVidia (The NV25 slightly improved the algorithm for Quincunx over the NV20, reducing the blurring and increasing edge AA quality).

 

[YeuEmMaiMai]

te problem is that FP16 does limit what you can do. Developers have to spend more time working around those limitations and therefore make adjustments to IQ as needed. Some examples:

Valve stated that you cannot do some of the lighting effects due to the fact that FP16 cannot handle the dynamic range and therefore will not be enabled on NV3X.

Now FP24 is superior IQ wise to FP16 just as FP32 is supperior to FP24 in all cases.


Now what if the positions were reversed with ATi and Nvidia? I would be owning an Nvidia card and not an ATI one....

Now JC also states that that NV30 path is inferior to ARB2 path in IQ
and he also states that ARB2 on NVidia is better than ATI but the performance is not acceptable.....

ALso JC stated that when you run ARB2 path R300 appears to be twice as fast as NV30 so I fail to see where you say what I stated above is not true........

And all this time I've thought that your gist was that using FP16 would result in a loss of image quality, which claim you still have not supported with quotes. I would assume that there isn't any in the S.T.A.L.K.E.R. game either (unless it's for Ati, if those rumors about FX-only shaders are true).

Yes, FX is slower than Radeon, but using FP16 keeps it at playable framerates and does not seem to result in loss of IQ. Notice that filtering tricks are a separate issue altogether and don't really have anything to do with shader precision.

 

[KimB]

te problem is that FP16 does limit what you can do.

No it doesn't. FP32 is there when FP16 is inadequate. It is in no way a limitation.

Valve stated that you cannot do some of the lighting effects due to the fact that FP16 cannot handle the dynamic range and therefore will not be enabled on NV3X.

Valve is wrong (that is, if you quoted them correctly, which is doubtful). FP16 has quite enough dynamic range for any sane lighting algorithm. FP16 supports a dynamic range of about 65,000:1. Can you honestly tell me that any lighting algorithm is going to require more? I don't think so.

Now JC also states that that NV30 path is inferior to ARB2 path in IQ
and he also states that ARB2 on NVidia is better than ATI but the performance is not acceptable.....

The NV30 path is sure to use quite a bit of FX12, which has significant limitations in dynamic range. DOOM3 has code that allows all the dynamic range that is needed when using such integer formats, so that there may be banding in places for the NV30 path. I would expect that there will be a path that changes all FX12 calls to FP16 calls for the NV35+.

When JC stated that he had code that allows the use of HDR on any video card, he also stated that the artists were not using that code at the time. Thus, we do not know what the actual quality differences will be.

 

[Doomtrooper]

te problem is that FP16 does limit what you can do.

No it doesn't. FP32 is there when FP16 is inadequate. It is in no way a limitation.

On NV35 chipsets, on NV30/NV31/NV34 chipsets the driver exposes only s10e5, or don't those cards exist that are being flooded into the market. Most DX9 lighting also uses FP render targets

 

[Neeyik]

te problem is that FP16 does limit what you can do.

No it doesn't. FP32 is there when FP16 is inadequate. It is in no way a limitation.

On NV35 chipsets, on NV30/NV31/NV34 chipsets the driver exposes only s10e5, or don't those cards exist that are being flooded into the market. Most DX9 lighting also uses FP render targets

Both of the 5200s that I'm "lucky" enough to be testing right now expose S23e8.

 

[Doomtrooper]

What driver revision is that Neeyik ?

 

[Neeyik]

52.16 I think....lemme check....yup!

 

[Doomtrooper]

Well from what I saw from this thread, Low and High precision react the same, which is what was exposed in the earlier 45 and 52 series pixel shader precision.

Maybe you can verify yourself.

http://www.beyond3d.com/forum/viewtopic.php?t=9525&postdays=0&postorder=asc&start=20

 

[Neeyik]

I'll download it now and test it out when the 5200 has finished on another batch of thingies I've got it chugging through. For now, you can chew through some Fill Rate Tester results:

Pure fill rate	561.8
Z pixel rate	256.4
Single texture	359.6
Dual texture	269.8
Triple texture	155.1
Quad texture	139.8
PS1.1 simple	245.3
PS1.4 simple	93.7
PS2.0 simple	155.7
PS2.0 simple PP	155.7
PS2.0 longer	93.3
PS2.0 longer PP	93.3
PS2.0 4 registers	76.2
PS2.0 4 registers PP	93.2
PS2.0 per pixel	13.8
PS2.0 per pixel PP	14.3

 

[Simon F]

A bit late but..

Crikey, there ain't much Christmas spirit and/or goodwill around, is there?

Christmas spirit? Bah! Rambug!

 

[Dio]

I can't believe you waded through all this tosh in order to read that!

Welcome back, hope you had a good Chrimbo (down under?).

 

[Simon F]

I can't believe you waded through all this tosh in order to read that!

Well it was early on... I lost interest at about page 300 or something ;-)

Welcome back, hope you had a good Chrimbo (down under?).

Cheers. Christmas in Blighty and New Year in Turkey. (Can't afford to fly to Aus' for Christmas... the fares are extortionate!)