R420 has 32-bit shaders

[rwolf]

http://www.beyond3d.com/forum/viewtopic.php?t=9273

Or at least that is what these guys are suggesting. The site they are referring to had lots of inaccurate info on the R3X0 chips though.

[Hanners]

Personally, I'd be hesistant to try and extrapolate any info from an interview after it's been translated a couple of times...

[The Baron]

however, it's been suggested before at B3D with the "Future Hardware Simulator" comment that was generating images that were almost identical to Refrast (<10 pixels or something). Uttar says it's not true, but hey, he says his ATI sources suck anyway. if it has FP32, I will laugh RIDICULOUSLY hard if it supports FP16 as _pp as well.

[digitalwanderer]

I don't think so. I don't know it for a fact, but I've heard nothing but 24FP talk out of all my ATi sources and they've stated that they didn't intend to make the jump to 32FP until dx10 when it would be required...which is still a ways away.

Sorry, I just do not see it and find the idea very uncreditable. If no one else does I'll ask about it at the R3D chat tomorrow. ;)

[Hanners]

Originally posted by digitalwanderer
If no one else does I'll ask about it at the R3D chat tomorrow. ;)

Like they'll say anything about it... ;)

[tEd]

Originally posted by Hanners
Like they'll say anything about it... ;)

they will:

"no comments on unannounced products" ;)

[StealthHawk]

Originally posted by The Baron
however, it's been suggested before at B3D with the "Future Hardware Simulator" comment that was generating images that were almost identical to Refrast (<10 pixels or something). Uttar says it's not true, but hey, he says his ATI sources suck anyway. if it has FP32, I will laugh RIDICULOUSLY hard if it supports FP16 as _pp as well.

There's nothing wrong with partial precision. The problem arises when full precision is too slow to be usable. If you can get FP16 support without too much time/work and it is even 10% faster than FP32 when many devs are apparently going to support partial precision anyway....why not support it?

PP support is basically going to be a given in DX9 games. If ATI can hop onboard the train and use FP16 without degraded quality(as I would hope FP16 is being used now) it would be kinda stupid not to support it. Let the competition get the ball rolling, and then catch a free ride.

[GlowStick]

seems more like a pritty open ended statement

/*Translated*/
Q > Will you ever change the persion you are useing now from this exact moment untill the end of time.

A > Um, sure why not.
/*Translated*/

[Hanners]

Originally posted by StealthHawk
There's nothing wrong with partial precision. The problem arises when full precision is too slow to be usable. If you can get FP16 support without too much time/work and it is even 10% faster than FP32 when many devs are apparently going to support partial precision anyway....why not support it?

PP support is basically going to be a given in DX9 games. If ATI can hop onboard the train and use FP16 without degraded quality(as I would hope FP16 is being used now) it would be kinda stupid not to support it. Let the competition get the ball rolling, and then catch a free ride.

The problem then is that you are at the mercy of the developer. Yes, using partial precision should be simple, but you have no guarantee that it will be used wisely, or correctly.

As an aside, I wonder if it would be possible for ATi to implement FP32, then use FP24 as the fallback mode for partial precision? Would it even be worth doing that, or would it be too expensive from a transistor point of view?

I imagine a lot of what ATi decide to do will depend on how fast a FP32 implementation they feel is possible - I can't imagine them wanting to use it unless they can squeeze as much performance as possible out of it.

[nobie]

Presumably, supporting FP32 and FP24 would cost even more transistors than supporting FP32 and FP16.

[StealthHawk]

Originally posted by Hanners
The problem then is that you are at the mercy of the developer. Yes, using partial precision should be simple, but you have no guarantee that it will be used wisely, or correctly.

True, but I think we have to agree that partial precision hints will be used in all major games using DX9 shaders. Does anyone actually know of any single game coming out using PS2.0 that doesn't also use _pp?

As an aside, I wonder if it would be possible for ATi to implement FP32, then use FP24 as the fallback mode for partial precision? Would it even be worth doing that, or would it be too expensive from a transistor point of view?

It would be a lot more expensive than implementng FP16 I imagine. Probably a reason why they don't have partial precision in R3xx- what would it have even been?

I imagine a lot of what ATi decide to do will depend on how fast a FP32 implementation they feel is possible - I can't imagine them wanting to use it unless they can squeeze as much performance as possible out of it.

Agreed.

[Evanescence]

I've think 32bit shader comes but not with R420. This costs two time more transitors than a normal 24 unity. If ATi really want to use 12x1 architecture, it's impossible to produce this chip.:eek:

Don't expect antil R500

[-=DVS=-]

Makeing 16pipe chip with only 16bit shader precision would make more sense , why do we need so much precision so early in DX9 development, need more speed :D more pollys and more fillrate.

[Blacklash]

Agree with Stealth here,

If I was running things, I would care less who started what. I would do what was best for the industry, consumer, and real performance of my product...

[Hanners]

Originally posted by -=DVS=-
Makeing 16pipe chip with only 16bit shader precision would make more sense

Not really, because then you'd be taking a step backwards from an image quality sense and you also wouldn't be following the minimum DirectX 9 spec. It really wouldn't be a good idea at all.

[Fiver]

The new maximum PC says that loki will have higher persision than r350, which leads me to believe that it will support 32 bit shaders.

[GlowStick]

Originally posted by Fiver
The new maximum PC says that loki will have higher persision than r350, which leads me to believe that it will support 32 bit shaders. Hm it could be ture then, unless the magazine was speculating from an obscure comment some ati rep made, but magazines usually dont do that stuff.

[Fiver]

That was my thought exactly. between that and some comments certain ati employees have made recently i'm pretty sure they will.

[Uttar]

As DaveB recently made me note...
R420 is not always the same thing as Loki.

Which means knowing what's true and what's not will be EXTREMELY difficult when it comes to Loki.


Uttar

[Fiver]

I thought r420 was code named Loki?

[nForceMan]

Originally posted by Hanners
Not really, because then you'd be taking a step backwards from an image quality sense and you also wouldn't be following the minimum DirectX 9 spec.

(omg) Why are you insisting on that :bs: ? Have you ever read DX9 spec at all?

[The Baron]

Originally posted by nForceMan
(omg) Why are you insisting on that :bs: ? Have you ever read DX9 spec at all?
FP24 is minimum precision for full precision shaders.

[nForceMan]

Originally posted by The Baron
FP24 is minimum precision for full precision shaders.

Source?

[The Baron]

Originally posted by nForceMan
Source?
NVIDIA, mother****er (can you tell I've been waiting to do this?) (http://developer.nvidia.com/docs/IO/8230/D3DTutorial1_Shaders.ppt)

(I blame Stealth for making me enjoy this so damn much. no offense to Stealth or nFM. :p )

[nForceMan]

Originally posted by The Baron
NVIDIA, mother****er (can you tell I've been waiting to do this?) (http://developer.nvidia.com/docs/IO/8230/D3DTutorial1_Shaders.ppt)

(I blame Stealth for making me enjoy this so damn much. no offense to Stealth or nFM. :p )

It does not carry conviction. It's just a presentation. You should quote the appropriate part of the DX9 spec. (popcorn)