David E. Orton's talk

[Tim Murray]

DX10 is not until Longhorn, or so the story goes.

It might support FP32, but it might not be exposed. Maybe just to give the hardware guys practice before R500. Then again, it would be strangely humorous if it supported FP32 and FP16.

 

[Hanners]

It might support FP32, but it might not be exposed. Maybe just to give the hardware guys practice before R500.

Surely that would be wasteful in the extreme, given the number of transistors FP32 support would most likely require?

 

[Tim Murray]

It might support FP32, but it might not be exposed. Maybe just to give the hardware guys practice before R500.

Surely that would be wasteful in the extreme, given the number of transistors FP32 support would most likely require?

Probably, but it would definitely keep them from screwing up on R500. But yes, you're right, it is wasteful and would only serve to increase the cost of the chip. I really have no idea what they're doing. I'd be surprised if they use FP32.

 

[vb]

It might support FP32, but it might not be exposed. Maybe just to give the hardware guys practice before R500.

Surely that would be wasteful in the extreme, given the number of transistors FP32 support would most likely require?

Probably, but it would definitely keep them from screwing up on R500. But yes, you're right, it is wasteful and would only serve to increase the cost of the chip. I really have no idea what they're doing. I'd be surprised if they use FP32.

I don't think there's much to screw up. it either does or it doesn't process at 32 bits. the result is stored at 32 bits even now. it is not like other "exotic features" like looping in PS, texturing in VS...

Of course you can really really really really really really really really underestimate the nr. of registers needed but they're not amateurs, are they?

 

[Geo]

I would be careful about that site. They had all sorts of quotes from ATI guys saying the R300 or R350 was going to be .13 micron. Their information turned out to be quite inacurate.

Does anyone recognize those photos of Orton? i.e. are they stock photos? The reason it matters, I would think, is it makes a difference in judging their credibility if those are actually photos of Orton taken during this interview.

 

[MuFu]

...given the number of transistors FP32 support would most likely require?

What, like 200 million? That Synopsys announcement a while back would seem to indicate that such a figure is likely (DFT timeframe, blah blah blah...).

I'm sure they wanted FP32 in R3x0 and the main reason they couldn't make it happen was because of process/die size targets. Having said that, in hindsight it seems like a very good choice indeed, as results using FP24 have been more than acceptable for this generation.

Consider that the high-end lifespan of R3x0 will be about 18 months and presume it'll be a couple of years until we see R500. Do ATi think FP24 will still be satisfactory in Fall 2005? Probably not.

MuFu.

 

[jvd]

...given the number of transistors FP32 support would most likely require?

What, like 200 million? That Synopsys announcement a while back would seem to indicate that such a figure is likely (DFT timeframe, blah blah blah...).

I'm sure they wanted FP32 in R3x0 and the main reason they couldn't make it happen was because of process/die size targets. Having said that, in hindsight it seems like a very good choice indeed, as results using FP24 have been more than acceptable for this generation.

Consider that the high-end lifespan of R3x0 will be about 18 months and presume it'll be a couple of years until we see R500. Do ATi think FP24 will still be satisfactory in Fall 2005? Probably not.

MuFu.

Sorry mufu why do you think we wont see the r500 till after fall of 2005 ? From what I understand we could be seeing it in winter of 2004 or spring of 2005 ?

 

[nobie]

DX10 is not until Longhorn, or so the story goes.

I don't understand why DX10 is so closely tied to Longhorn. AFAIK the requirements for Longhorn are simply a DX9 feature set for the fancy interface, and DX7 for the trimmed down interface.

 

[KimB]

I don't understand why DX10 is so closely tied to Longhorn.

Because that's the way Microsoft wants it.

 

[Super Grafx]

R500 will come at least 18 months AFTER R420.

since R420 is coming spring 2004, of course we wont see R500 until fall 2005 at the soonest. factoring in a delay, it could be winter 2005/2006, or even spring 2006.

 

[Pete]

Yes, ATi mentioned their 18 month double-the-high-end-speed cycle when the 9700P was released. The question is whether they can maintain the rather heady momentum they now enjoy.

 

[jvd]

R500 will come at least 18 months AFTER R420.

since R420 is coming spring 2004, of course we wont see R500 until fall 2005 at the soonest. factoring in a delay, it could be winter 2005/2006, or even spring 2006.

ah. Unless ati themselves don't really consider the r420 a new highend chip and just consider it another version of the r300 ?

 

[bigbud120]

ah. Unless ati themselves don't really consider the r420 a new highend chip and just consider it another version of the r300 ?

I hope this is the case. It's a let down that the r400 had to be pushed into the r500. I feel like the r420 is just a filler until the r500 is finished.

Here's a link about 32-bit precision that I saw on Xbit today.

http://www.xbitlabs.com/news/video/display/20031202110836.html

 

[Dave Baumann]

I still bet on FP24 for now. PS3.0 is, according to my information, still FP24 for its PS requirements.

 

[991060]

I still bet on FP24 for now. PS3.0 is, according to my information, still FP24 for its PS requirements.

Does the information come from talks with MS DXDEV team? And will the "_pp" hint be reserved in PS3.0? I don't see it in the SDK document.

 

[Geo]

ah. Unless ati themselves don't really consider the r420 a new highend chip and just consider it another version of the r300 ?

I hope this is the case. It's a let down that the r400 had to be pushed into the r500. I feel like the r420 is just a filler until the r500 is finished.

If Loki was R390 that argument would make more sense. Calling it R420 indicates the opposite.

 

[Joe DeFuria]

If Loki was R390 that argument would make more sense. Calling it R420 indicates the opposite.

Loki=R390=R420, IIRC.

 

[Geo]

If Loki was R390 that argument would make more sense. Calling it R420 indicates the opposite.

Loki=R390=R420, IIRC.

Yes, exactly. But that they settled on "R420" instead of "R390" is what I find to be Significant. ATI usually plays marketing games with board names instead of chip names, so until better info shows up eventually, I'm betting on "R420" *as a name* means they see it as a generational jump, even if not as ambitious as they had planned at one time for the R3xx to R4xx move.

 

[Dave Baumann]

I think thats wrong actally, since it was based on the assumption that there was only one R420. IMO...

R390 = (non-public) R420
Loki = What will be R420

The non public R420 is PS/VS2.0 AFAIK.

 

[Super Grafx]

the R400 / R390 / R420 / Loki is sure as heck one of the more confusing graphics processor(s) to keep track of.