r420 may beat nv40 in doom3 with anti-aliasing

See Baron? THAT'S why they don't ban him.

 

sigh, so true...

Beyond3D Drinking Game: Every time Radar says something stupid, take a shot. If you can get through a single thread without your liver exploding, you win.

 

I win!

 

R3xx/R4xx do not do everything at FP24 precision or else they would have no need of mini-ALU's that aren't FP24 capable.

 

Seriously, are you just making this stuff up to keep the conversation going?

 

The mini-ALUs are FP24 capable. You must be thinking of NVIDIA's parts.

-FUDie

 

"Mini" refers the the instruction set that those ALUs support.

 

You've got to be just putting us on now, right?

 

Basically the only thing that's done at FP32 in the R3xx's pixel shader is fixed-function texture addressing.

 

Last I recall, ATI has stated in interviews that all precision in the pixel shader is handled internally at FP32.. the output that results from it is FP24.

I presume it is a difference between the storage and the internal calculations.

 

Last I checked, NVidia ENDORSES the use of FP16 AND PS 1.4.

I also recall, ATI processes _PP too... which still runs better than NVidia's hardware (in the case of R3XX vs NV3X).

Of course, there's absolutely no correleration with this statement:

Who wanted _PP? NVidia
Which IHV with DX9 hardware doesn't support _PP? noone that anyone knows of
Whose hardware runs terrible w/o _PP? NVidia (at least with the NV3X hardware)
Which company's DX9 app (NOT BENCHMARK) hasn't used _PP? none that I know of

Flawed arguments with erroneus statements don't make what one says true.

 

You've got it backwards. The input/output formats supported in the R3xx go up to FP32. The internal pixel shader math is all done at FP24. Even if the actual units had more precision than that, they're still storing at FP24 in the registers after each operation, which would destroy the point of having higher precision math in the first place.

 

Okay, I got the FP24 bit wrong for the mini-ALU's. For that I apologise.

What I'd like to know is are the Mini-ALU's more suited to working PS1.x instructions or SM2.0 instructions?

Once again I'll point out that nVidia also did this with NV30, NV31, NV34 but when DX9 changed to full/partial precision from multi-precision they altered their mini-ALU's to reflect that. Unfortunately the change happened late enough that the above gpu's couldn't be modified in time (production chain already in motion) and first appeared in NV35. This more than anything else probably explains why NV30 was aborted from an architectural standpoint (not to say there weren't other reasons also, not related to architecture).

 

The best I can get from ATI is that they contain more than just PS1.4 modifiers. If they do contain other instructions then these can be scheduled to assist the full ALU for any other operation that requires those instructions.

DX9 was always floating point precision - there was no change. It is an implemantation detail how they support legacy integer shaders - IHV's had the option to support integer shaders with separate units or float units with conversion. NVIDIA gambled with NV30 that DX8 would still be the main requirement and so they kept a similar structure to NV25 with NV30 for the integer processing but extending the texture address processor (which is already float in NV2x) to support DX9 float operations. Id's say their issues stemmed from NV30's lateness and some blindsiding of ATI's architecture whereby ATI decided to make everything float which offered comparable, or better, DX8 performance and very good DX9 performance.

 

From what I've read about the NV40, perhaps the largest problem with the NV3x is that the compiler is the last thing they developed. This basically allowed the hardware team to make a lot of decisions that made life extremely hard on the compiler writers. This is why nVidia's drivers are tailor-made to specific shaders in games, and why the general compiler has been so slow to improve with the NV3x.

Of course, there's also the fact that there just wasn't much floating point performance to be had in the NV30-34 to begin with. This really should be the most surprising thing, if you think about it. Why would nVidia want to shoot themselves in the foot by automatically halving (or more) PS 2.x performance over PS 1.x performance? I can only conclude that early in development, nVidia planned on integer precision being allowed in PS 2.x.

 

I'm just happy that the NV40 parts are back "up to par" with the competition. Yeah, there's a bigger feature set without argument, but the performance in current situations is what's very nice.

For the past 18-24 months, the only NV hardware I've been able to wholeheartedly recommend were the NV25 series for great budget cards and the 5900NU series for good performance at a cheap price. Every other price/performance point was better covered by ATI, which was quite sad because it made the whole market lopsided.

This round is very obviously different in my eye; enough so that I am strongly considering ditching my X800XTPE and looking for a 6800U (although the prices are still a little insane on the NV40 right now).

 

I don't drink and I think my liver STILL exploded! Is that a bad sign?

 

Exploding liver is always a bad sign.

 

By integer precision you mean FX12 (48-bit)? Because that's what John Carmack was requesting for Doom 3. And NVIDIA usually listenes to him as they have the best OpenGL impementation in gaming cards. If I remember correctly the request for 48-bit color came as soon as in GF256/GF2 days... I think NVIDIA thought if FX12 is good enough for Carmack, it would be for Microsoft too, but time went on, they had a fight over XBox grf. chip pricing, Doom3 was nowhere to be found and NV3x was the one, that ended up Doomed...

 

I don't think your argument makes sense. For one, Carmack wanted higher precision backbuffers and FX12 only applies to the shaders. Second, I believe he was talking about 64-bit (FX16 per component) surfaces. NVIDIA's FX12 shaders were a byproduct of carrying over legacy shader stuff from the GF4.

-FUDie