Interview with ATI's CEO, Dave Orton

WOW

incredibly interesting interview.

I read it all, but I have not read this entire thread yet.

my first question is, and I probably already know the answer, regards this:

the Orlando development site, that would be where some of the former Lockheed Martin Real3D people are, right? because ATI got some of Real3D. (I've asked about ATI owning some of Real3D before, but this is the first official comment I've read from Dave Orton about the Orlando development site.

so R600 is a joint-effort from all 3 of ATI's design teams, although as he says, the architectural centre team is in the Valley (Santa Clara, ATI West).

soooooo, Xbox 2 will be designed by ATI, ArtX, Real3D teams with their respective technology all combined into one VPU! I'm just assuming so based on R600 being a joint effort between all ATI teams.

if true, coolness 8)

 

Anyone knows which team is working in N5 and which is in XB2 ?

PS:I supose not but with a litle of hope...

 

Okay, so I'll do my DW imitation:

I'm too thick to follow y'all, so throw a dog a bone. Is the *original* R400 now:

1). "He's dead, Jim". Tho possibly little bits 'n pieces will get recycled as R&D is never really wasted.

2). The basis for Xbox2 GPU.

3). The basis for some other ATI GPU, and if you say 'Yes', you have to say which one.

 

yep. but it seems to be ATI's style to talk about future generations long before they come out. ATI / Dave Orton mentioned R400 and R500 in May of 2002.


I must say this is by far the best and most up-frony interview I've ever read from any graphics company. Dave Orton is the man 8) especially since his company will lead the technology fight against Sony in the console arena.

 

I think ATI will support SM 3.0 - it will be in R500, which should be out by spring or summer 2005.

R600 and Xbox 2 VPU should be SM 4.0 and neither will be out until 2006.
that would be good. if Xbox 2 were to launch in late 2005, it would probably not be able to have DX10/SM 4.0 technology, which it really needs since it'll have to last until 2009-2010.

on the other hand, ATI commented last year that both the new Xbox and new Nintendo consoles would have DX9-level shaders, or at least DX9-level shaders.



now, my comments above do not really take into account the new Dave Orton interview. so everything may have changed. their might not be an R500 at all, etc etc.

 

Who's bashing? I was simply pointing out why what ATi says with regard to R4x0 presently supporting some, but not all ps3.0 functionality in r4x0, makes fairly good sense.

As far as developers go, what've they done with ps2.0 since R300 shipped 20 months ago? Little to nothing, right? So bashing ATi for stating valid reasons for R4x0 not supporting all of the ps3.0 feature set, but supporting several critical functions of ps3.0 nonetheless, seems pretty naive to me.

Here's a hypothetical for you of the type Orton was speaking of...suppose nVidia's nV40 chip yields are very poor because of additional circuitry, possibly transistors nVidia designed in to support some features of ps3.0 (those features ATi deliberately declined to support), and nVidia finds it has a real problem shipping a profitable number of nV40's to its AIB OEMs? Obviously, in that case, ps3.0 support in nV4x won't mean a hill of beans as far as what developers are able to support, will it, since very few people will be able to own nV40? Or, suppose that like ps2.0 support in nV3x, nVidia's initial support of ps3.0 in nV40, specifically the ps3.0 functionality ATi omitted from R4x0, is similarly as poor an implementation as nVidia did with ps2.0 in nV3x? How many developers do you suppose will be interested in investing the time and money to support nV40 ps3.0 in either or both of those events? My guess would be fewer than currently support ps2.0.

That's the sort of problem I think Orton was alluding to ATi avoiding by choosing the layout for R4x0 functionality. Would have been the same thing exactly had ATi designed the best ps2.0 support on earth but been unable to get yields sufficient to sell enough of them to motivate developer support. Ironically, ATi's shipped literally millions of competent ps2.0 supporting R3x0's into the market, and as I say that has attracted scant ps2.0 support from developers. So just why might we assume that at this point in time, with nVidia having shipped 0 ps3.0 gpus into the market, that ps3.0 support from the same developers would be any better?

Additionally, we ought not forget that for all of last year nVidia literally paid developers to eschew ps2.0 and stick to ps1.x. First thing nVidia's going to have to do assuming nV40 yields are acceptable and nV40's ps3.0 implementation (specifically the ps3.0 functionality excluded by R4x0) is also acceptable, is overcome the negative developer inertia nVidia was pushing so hard for all of last year as to *any* psX.x support beyond 1.x. I think at this stage it would be naive to describe the situation as being as simple as you represent for all of these reasons.

 

So what do we know now:

DB has indicated that ATI's next VPU will also be some form or other of R300 derivative.

The leaked slides indicate that ATI will at some point have an SM3.0 part.

The questions that then leap to my mind is just what eventual limits are they going to hit with R300 derivatives, and what qualities define a card as an R300 derivative. Is it shader model support or something else? In other words, can ATI produce an R300 derivative with SM3.0 support, or is that something that is precluded by the architecture?

My own belief is that they can produce an SM3.0 compliant R300 based card, but that's based on what I believe to be the design constraints the R300 architecture is going to force on them.

My own belief is that they are saying their next card (after R42x) will take the following key features from the original R300 design: It will be a single precision part; it will have the same basic functional units and they will be laid out in the same basic organization on the die. It will also be organized in quads.

FP32 support won't be a huge problem, as it has been repeatedly stated that the main reason that R300 supported FP24 rather than FP32 is that the spec called for FP24. It's been further stated that an FP32 R300 derivative would have the same performance as an FP24 variant, but would need a larger die for the extra transistors needed to support FP32.

So the only real question is can the features needed to support SM3.0 be added to R3xx based shader hardware, and if not can supporting shader cores be dropped into those positions on the die. If so then ATI can do it.

That's my thought.

 

Ironically, at the time for the SM 2.0 spec, nVidia was pushing a specification that was much more feature-rich than the ATi one. Had Microsoft decided to take the nVidia specs, it would have been the same thing as it is now with SM 3.0.

Ok, nVidia lost, ATi got their specs approved as SM 2.0, but not much has changed. The disparities are the same this time around, just the official status has shifted from the one to the other.

 

I think the questions might be: can a designd based upon R3xx do SM 3.0 very well? And would it earn ATi a nice profit to try and design such a part?

 

Agree

 

So, basically, you're saying that it's M$'s fault nV3x was comparatively so feature-poor compared to R3x0? Remember that fp32 support in nV3x was useless for 3d-gaming, and I don't know how you might consider fp16 more "feature rich" than fp24, and who cares at all about the unused and discarded FX12? And then there are several API features relative to DX9 that R3x0 supports but nV3x doesn't support at all. What's your theory as to the features nVidia was prevented by M$ from implementing in nV3x? (I'd like to point out that it seems that had nVidia actually planned a much more "feature-rich" nV3x than they actually produced, they could've supported that richness under OpenGL extensions whether M$ supported them under DX9 or no, right? They could've hawked such features like they did "Ultra Shadow," right? I'm not one for blaming nVidia's troubles on M$.)

Which begs the question of why nVidia "lost," doesn't it? Could it be that nVidia "lost" because M$ didn't see all of the "feature richness" that you apparently think lies latent and hidden in nV3x? I certainly think that's a fair bet.

Also, nV3x does indeed support ps2.0--the problem was that the nV3x implementation of ps2.0 wasn't much contrasted with the R3x0 implementation of ps2.0. So I can't see how ps2.0 was "ATi's" specification any more than ps1.x can be said to be "nVidia's" specification, and I surely don't see how ps3.0 belongs to nVidia. Rather, it seems plain to me that all are D3d specifications that "belong" to the API, right?

My thinking about nVidia's ps3.0 support, providing nV40 yields are good and the nV40 ps3.0 implementation as a whole is good, is "more power to 'em".... I wouldn't be able, nor would I try, to find fault with nVidia under those conditions, certainly. My thoughts with respect to ATi and its partial support of ps3.0 functionality, described as ps2.a/b +, is that I see no gross fault in their reasoning, regardless of nV40 yields and the efficacy of nV40 ps3.0 support. I have to point out, though, that as of right now we have no clear picture of nV40 yields, nor do we know whether the nV40 implementation of the ps3.0 functionality ATi doesn't support will be compelling. So for me the issue is still up in the air.

 

Methinks it is number two... but it will be upgraded as process technology will improve. Malborough are in charge but of course get input from the Valley (the head of tech it seems). However since it is based on the R400 but upgraded from a performance (and not featureset perspective, that must be set in stone by now and I would think it is at above SM3.0 but not quote SM4.0) perspective it is internally known as R640*480 (see what they did there, quite clever eh?!).

Anyways no one talked about Nintendo's new part and me thinks that is another special project for Malborough (18-24 months sounds about the right timeframe for both XB2 and GC2).

About the PC Parts my feeling from that interview is that the R500 was glossed over and more importance was placed on R600. R500 could be SM3.0... for 2005 and 2006 = DX Next.

So perhaps we have a situation where R300 - R420 - R500 are all incremental evolutionary steps all having DX9.0x as the backbone.

 

Dave is one of the most personable people I have ever met (Err by Dave I mean both Dave Baumann and Orton are personable people). He's always been straightforward and open in the few conversations I've had with him. It's really too bad I couldn't publish the entire conversation with him from the R300 launch 2 years ago, as much of what we discussed happened as he put forth (Darn ATI PR )

 

Exactly. Makes you wonder what had happened if the nVidia specs would have been approved 2.0, doesn't it? I guess that would have made the ATi specs 1.8 or such, with some 2.0+ extensions...

(Btw. I'm not blaming Microsoft for the bad NV3x performance, obviously.)

Those DX specs are for the most part a moving target, filled with a lot of nice to haves. Some are arguably more important for making great looking games than others. Compromises, compromises.

I actually think that SM 2.0 was more or less the collection of features both ATi and nVidia expected to have.

Right. But I would assume, that 3.0 was the compromise of what Microsoft wanted and nVidia expected to deliver in their future part, the NV4x.

Agreed.

I would like to add that, following this train of thougt, that SM 4.0 probably is "whatever is in the ATi chip at that time", most likely, "and the competition thinks it can deliver" as well.

 

ATi is on record as saying if DX9 was fp32 required, r300 would have been fp32. I guess the difference would have been the R300's would have had lower margins and prices would have remained higher.

I'd say consumers should be glad that fp24 was the accepted spec.

 

It just came into mind why Dave O. or some other ATI guy said 2 years ago that R300 would be "revolutionary" and the next parts "evolutionary", R420 and R500 being based on R300 would finally explain this

 

Yes. That ATi was contributing willingly to the design of the specs and was prepared to change their chip to fit them, while nVidia stormed out and shouted foul, might have had some impact on the specifications agreed upon, wouldn't you agree?

That probably made it quite acceptable to Microsoft to accept FP24 over FP32, as it would be good for ATi and made no real difference anyway. That is even disregarding the possibility, that nVidia was the only one that wanted FP32 in the first place.

 

Don't you think that tibit is directly related to Xbox 2 and gives us a good indication about the timeframe ?

 

Why did he not mention the R420 architecture? Was it developed at a different site, or does he consider R420 a R300 derivitive and R500 something new?

No mention of R700.

Perhaps clean sheet designs will happen every two product generations?

Conversley, let's look at the following two comments:

While we don't know what "that" is, one reasonable assumption would be the R400 product that was in development.

The second comment being from DaveB that the XVPU would be R600 based.

These two comments when put together could read that the R400, rather than becoming the R500, became the R600?

DaveB's thoughts that that the R500 was something else entirely than what he thought it was would back this up. R500 being another product based upon the R300 architecture possibly?

Plenty of food for thought anyway.

 

Oops, forgot to login.