The Rumor Roundup

[Tim Murray]

Well I think the next DirectX Pixel Shader model does require 32-bit shader precision.

By next, you mean PS 3.0 or post PS 3.0?

I meant Pixel Shader 3.0 model (with roughly 65356 instructions).

Nope, PS3.0 increases register requirements but does not move the minimum full precision to FP32.

 

[Joe DeFuria]

Dr. DeFurious, I said R420 will have very high memory clocks for a reason. I don't know for sure, seeing as how I don't have a board, but if R420 does have ultra-high memory clocks, I won't say I told you so. (There are things I'm not telling you. Leave it at that. This looks to be the one area where ATI will be *very* aggressive.)

I truly hope you are right. I tend to put a weighty emphasis on historical precedent (which of course, does not always pan out.) So far, this strategyhas panned out wrt to PowerVR delivering a bleeding edge PC part. 8)

Oh, and You can say you told me so, I don't mind.

 

[Evildeus]

R420 will have FP32 support.
FP24 couldn't realistically be _pp, as it wouldn't be feasible to gain a performance benefit out of having both FP24 and FP32 support.

Could Ati do the same thing as Nv? Meaning, 3 FP24 to make 2 FP32? Then the 8 extreme or 12 pipilines makes sense. 8)

Edit: Joe made it faster than me

 

[volt]

Well I think the next DirectX Pixel Shader model does require 32-bit shader precision.

By next, you mean PS 3.0 or post PS 3.0?

I meant Pixel Shader 3.0 model (with roughly 65356 instructions).

Nope, PS3.0 increases register requirements but does not move the minimum full precision to FP32.

That might be true for current DX version though.

 

[Tim Murray]

Volty, no current card has anywhere near the requirements for whatever DXNext will require (which hasn't been finalized yet). So, PS3.0 is as far as we can go.

 

[Joe DeFuria]

Nope, PS3.0 increases register requirements but does not move the minimum full precision to FP32.

That's what I thouhgt.

On a related note, a while back I posed the question: if having the PS in fp24 would cause "interoperability" with VS working in fp32 mode. That is, we know that VS 2.0 can work with PS 3.0....but would that get ugly if the VS is in 32 bit if the PS is only 24?

I was looking for an answer to that question (which I don't think I ever got), when trying to decide for myself the liklihood of R420's PS units supporting fp32.

 

[Tim Murray]

Nope, PS3.0 increases register requirements but does not move the minimum full precision to FP32.

That's what I thouhgt.

On a related note, a while back I posed the question: if having the PS in fp24 would cause "interoperability" with VS working in fp32 mode. That is, we know that VS 2.0 can work with PS 3.0....but would that get ugly if the VS is in 32 bit if the PS is only 24?

I was looking for an answer to that question (which I don't think I ever got), when trying to decide for myself the liklihood of R420's PS units supporting fp32.

For the most part, I'm going to guess that the two operate independently. I don't think that there would be any problem with FP32 VS3.0 units and FP24 PS2.0 units. However, I really don't know--this is, after all, a guess.

 

[volt]

Cool Tim, my gripe with your post has been solved 8)

 

[991060]

On a related note, a while back I posed the question: if having the PS in fp24 would cause "interoperability" with VS working in fp32 mode. That is, we know that VS 2.0 can work with PS 3.0....but would that get ugly if the VS is in 32 bit if the PS is only 24?

I think VS has been working in FP32 since DX8 generation hardware. There shouldn't be a problem.

 

[digitalwanderer]

Just since it applies to the topic, here's Hanners' "Piecing together the puzzle - NV40 and R420/423 investigated" story he just put up that I told ya about that I thought he was going to kill ya over.

He's not, he wouldn't even call you a sea-slug in the story even though I asked him to many times!

Could Ati do the same thing as Nv? Meaning, 3 FP24 to make 2 FP32? Then the 8 extreme or 12 pipilines makes sense.

Isn't that what MuFu was asking me if anyone else had mentioned to me?

 

[StealthHawk]

I meant Pixel Shader 3.0 model (with roughly 65356 instructions).

EDIT: and next most likely DX9.1 (it'd be nice if anyone can confirm this). Funny thing PS3 is exposed within DX9 and it does not require high FP.

What makes you think that PS3.0 requires FP32 then?

 

[Evildeus]

I don't know. But that the only (stupid?) answer that comes to my mind at the moment
BTW:

What matters most to me is, that even the ati aibs said, that the nv40 will be faster this time.

Moreover my soruces said, that the r420 and 423 will be "just" a wider r300. So, ati did not change much. Whereas nvidia did a whole redesign.

http://www.nvnews.net/vbulletin/showthread.php?t=26440

 

[volt]

I meant Pixel Shader 3.0 model (with roughly 65356 instructions).

EDIT: and next most likely DX9.1 (it'd be nice if anyone can confirm this). Funny thing PS3 is exposed within DX9 and it does not require high FP.

What makes you think that PS3.0 requires FP32 then?

One of the roadmaps I've seen. Nothing special I guess, though it was worth to clear things up.

 

[Tim Murray]

PS3.0 does not nor will have anything to do with FP32. FP24 is still minimum required precision, FP16 is still minimum partial precision. Period end of story quack.

 

[ninelven]

Could Ati do the same thing as Nv? Meaning, 3 FP24 to make 2 FP32? Then the 8 extreme or 12 pipilines makes sense.

To my knowledge, this is rather impossible sans some rather remarkable technology.

 

[Joe DeFuria]

I think VS has been working in FP32 since DX8 generation hardware. There shouldn't be a problem.

Yes, I know that at least in VS 2.0, the vertex shaders are already FP32.

But my question is, since (I believe) VS 3.0 can now operate on textures...just like Pixel shaders, there is more "interoperability" between VS and PS with version 3.0 In other words, version 3.0 is sort of the first step to "unifying the shaders", which is different than in the past where they were pretty much separate entities.

So if the hardware has differences in underlying precision, I don't know if this impacts the usefullness of taking advantage of the new interoperability between VS and PS in 3.0.

 

[Tim Murray]

I don't think VS3.0's ability to read textures will really have any effect on pixel shaders, or at least in the way you're worried about.

 

[KimB]

Could Ati do the same thing as Nv? Meaning, 3 FP24 to make 2 FP32? Then the 8 extreme or 12 pipilines makes sense. 8)

Edit: Joe made it faster than me

I just don't think it's feasible. I'm not even sure anymore that the NV3x does execute more FP16 ops than FP32 ops (except for RSQ). It appears that FP32 performance would be just as high as FP16 performance on the NV3x if it weren't for the FP register limitations.

And taking three FP24 pipelines to make two FP32 pipelines would be vastly more complex in terms of transistors. I just really don't think so.

 

[Rugor]

I'm going to make one prediction for R42x -- It will be single-precision.

They may or may not go to FP32 from FP24, but I really don't see as major a change as a multi-precision architecture, especially since in DX9 the only other real option is FP16 for partial precision, and I can't see ATI bothering with it.

As to PS3.0, not guaranteed but after the way ATI blindsided everyone with R300 I'm not going to go out on a limb either way.

 

[Rys]

I guess we can probably confirm some product names too.

FX6000 and R9900 anyone?

http://www.gzeasy.com/newsphoto/y2k4/03/24/ati.jpg
http://www.gzeasy.com/newsphoto/y2k4/03/24/NVIDIA.jpg

Rys