Inquirer spreading R420 info

[Seiko]

One more comment.

As for my expectations, yes I know I'm asking a lot infact I'd even go so far as to say I'm being unrealistic at this current moment in time. I appreciate the hardware vendors are finding it difficult but have we really come that far in the 5 years or so I've been following this industry?
It's probably just my lack of patience more than anything else but the sooner the base line platform shifts to DX9 the sooner we can expect to see the next level of programs and applications.

Let's look at history a little. The TNT was released in 1998 if memory serves. The peak fillrate of the TNT was 160 million pixels per second. These pixels were of DX6 caliber (i.e. limited texture blending modes.) Now it's 2004 and the NV40 is here. The peak fillrate of the NV40 is 6400 million (colored) pixels per second. These pixels can be shaded with PS 3.0 programs in full 32-bit floating point. So we've come 35x in raw performance in 6 years and have much more flexibility to boot. Now can you can see why I called your viewpoints "childish"? Other vendors have had similar leaps in performance and quality (anyone recall the performance of ATI chips in the 1998 timeframe?).

Go ahead and whine about how "companies don't care about consumers and things haven't advanced fast enough", it's amusing. Maybe you should complain about how games aren't taking enough advantage of these new features. UT2004, for example, is still largely a DX7 game.

-FUDie

Although my comments could be assessed as very demanding, unrealistic even I don't think that equates to childish. The fact of the matter is as you say here we are in 2004 and the cutting edge card still only offers 4xFSAA, an improvement over 1998 hardware yes. An improvement over my old 9700Pros quality, hmmm, I'd have to check really but honestly suspect it won't be. Ok, some will be thrilled at it's raw performance and additional features. As I said though as a very demanding consumer I wanted it to have features, performance and IQ improvements. So I got 2 out of 3 with the NV40, some will say that's not bad going. Personally, and again only in my very high standards it missed the mark.
Now extrapolate when we'll actually see this cards performance as the baseline, 2006 maybe? That's a long way away so you'd better be absolutley thrilled with it!

 

[jpaana]

Ati exposed 12 registers out of 32 avaliable. I remember sireric talking about it

If 32 were actually available, they'd be exposed by now.

2) The R300 has 32 temporary registers in the vertex and pixel shaders (64 "total"). We currently "reveal" 12 in the pixel shader (not sure about vertex shader), following DX9 recommendations. We will raise that as caps bits allow or DX9 specs change.

This would make the 2.0b pretty ideal for R350+ at least and possibly usable on R300 as well as according to the Richard Huddy presentation video from GDC (unfortunately no juicy powerpoint bits there) they are doing pretty well in collapsing multiple shader assembly instructions to less hardware instructions (from around 90 to 45 instructions, it's between 11:30 and 12:00 in the video, pretty optimistic but anyway) and thus could run longer shaders than the current limit allows (as runtime, not driver, enforces it).

 

[PaulS]

Although my comments could be assessed as very demanding, unrealistic even I don't think that equates to childish. The fact of the matter is as you say here we are in 2004 and the cutting edge card still only offers 4xFSAA, an improvement over 1998 hardware yes. An improvement over my old 9700Pros quality, hmmm, I'd have to check really but honestly suspect it won't be. Ok, some will be thrilled at it's raw performance and additional features. As I said though as a very demanding consumer I wanted it to have features, performance and IQ improvements. So I got 2 out of 3 with the NV40, some will say that's not bad going. Personally, and again only in my very high standards it missed the mark.
Now extrapolate when we'll actually see this cards performance as the baseline, 2006 maybe? That's a long way away so you'd better be absolutley thrilled with it!

The 9700 was only released September 2002, and we're still in the first half of 2004, so of course we won't have come that far in terms of AA quality. Having said that, the R420 will please you if that's what you're after.

But yes, you are being pretty unrealistic about the kind of increases you expect. I'd say both ATi and NVIDIA are progressing at a ridiculously fast rate right now - at least 2x the performance of the last gen isn't to be sniffed at. If you want to point the finger at anyone, it's the CPU guys who possibly aren't doing all they can at the moment (not that I blame them, particularly). Really wish they'd get those multi-core chips out, so CPU performance can shift up a gear again.

 

[Natoma]

I doubt we'll see multi-core cpus in the consumer space for at least another 2-3 years. Intel can't even get single core 0.09 micron process working right, let alone a dual core package.

Of course, they could go the pentium pro route and separate the cores rather than putting them on the same die. But then it'd be hideously expensive like the P4 EE, or worse.

 

[Seiko]

But yes, you are being pretty unrealistic about the kind of increases you expect. I'd say both ATi and NVIDIA are progressing at a ridiculously fast rate right now - at least 2x the performance of the last gen isn't to be sniffed at.

I'd agree with you Paul to some extent. ATI showed what they could do with the R300. Simply stunning, the card from my perspective at least changed the landscape forever. The 9500Pro was a nice(ish) improvement over the mid range cards at the time when FSAA and AF where brought into play. Then to my horror ATI released a rebadged 8500 into the low end as opposed to a genuine DX9 card to clear out old chips? One corner cut I thought! Then came the 9600, a chip designed from what I could gather at least to increase yields and increase margins. I guess as some have suggested this could have given ATI the ability to play with new techniques and I may soon reap the benefit but as a consumer it didn't sit right due it's slower performance compared with the 9500Pro and again no new features. Fast forward a little and then came the R350. Hmmm, no new features that I could really see, just a tiny tweaked core, a Mhz bump and nice new box. Looking at the performance delta between the R300 and R350 it was very evident that ATI had at least from the surface done very little.

Anyway, as you said, I may be expecting far too much and looking at the top range cards I think I'd agree ATI and Nvidia have shown what they really can do. It's just a shame their mid range and low level parts don't have the same bang.

Finally, don't think I didn't notice that teeze about ATIs new FSAA. I just hope you're right!

 

[Hellbinder]

Finally, don't think I didn't notice that teeze about ATIs new FSAA. I just hope you're right!

He is right

 

[Randell]

GF3 - Gf3Ti500/Gf3Ti200 - which is exactle the same as 9700Pro - 9800Pro in execution. An extra refresh in the XT was added because of the lengthening product cycle.

Except I would argue that the ti500 didn't differ much at all from the original GF3, aside from higher clocks to settle the 8500/GF3 debate for the next couple of months.

realistically what did 9800pro offer consumers other than clock/mem speed? I'm saying they're similar examples of the product/refresh cycle. So I think we agree!

 

[KimB]

Ati exposed 12 registers out of 32 avaliable. I remember sireric talking about it

If 32 were actually available, they'd be exposed by now.

2) The R300 has 32 temporary registers in the vertex and pixel shaders (64 "total"). We currently "reveal" 12 in the pixel shader (not sure about vertex shader), following DX9 recommendations. We will raise that as caps bits allow or DX9 specs change.

The DX9 specs have always allowed 32 temporary registers. The necessary caps bits have been there since introduction. DX9 was never a reason not to expose these.

C'mon, they've had a year and a half to expose those 32 temporary registers.

 

[AlphaWolf]

Are you insinuating that sireric was lying about it?

Strangely enough I am more inclined to believe him than I am you.

 

[OpenGL guy]

Ati exposed 12 registers out of 32 avaliable. I remember sireric talking about it

If 32 were actually available, they'd be exposed by now.

2) The R300 has 32 temporary registers in the vertex and pixel shaders (64 "total"). We currently "reveal" 12 in the pixel shader (not sure about vertex shader), following DX9 recommendations. We will raise that as caps bits allow or DX9 specs change.

The DX9 specs have always allowed 32 temporary registers. The necessary caps bits have been there since introduction. DX9 was never a reason not to expose these.

C'mon, they've had a year and a half to expose those 32 temporary registers.

Listen to what people are saying, you might learn something.

12 temps is what you get when exposing PS 2.0. If you want to expose more than 12 temps, then you have to go to PS 2.x. What would be the point of enabling PS 2.x just to expose more temps? Microsoft said that 12 was enough for PS 2.0 and the HLSL compiler is built around that.

The HW has 32 temps, realistically we could have exposed 16 for R300, but since it requires changing to PS 2.x, what's the point? Since everyone is writing PS 2.0 shaders with 12 temps in mind (i.e. so they work on all platforms), again what would be the point of exposing more temps? Are people saying "my temps are bigger than your temps"?

PS 2.0 allows for 64 ALU instructions, can you really see a need for 32 temps? 12 should be plenty.

 

[tEd]

Ati exposed 12 registers out of 32 avaliable. I remember sireric talking about it

If 32 were actually available, they'd be exposed by now.

2) The R300 has 32 temporary registers in the vertex and pixel shaders (64 "total"). We currently "reveal" 12 in the pixel shader (not sure about vertex shader), following DX9 recommendations. We will raise that as caps bits allow or DX9 specs change.

The DX9 specs have always allowed 32 temporary registers. The necessary caps bits have been there since introduction. DX9 was never a reason not to expose these.

C'mon, they've had a year and a half to expose those 32 temporary registers.

i think they could have exposed 32temps for the compiler target but it would have need a seperate compiler target besides the ps2.0 and i think it would not be worth it just for a different temp count.

 

[Nick Spolec]

You tell him, OpenGL Guy.

 

[KimB]

12 temps is what you get when exposing PS 2.0. If you want to expose more than 12 temps, then you have to go to PS 2.x. What would be the point of enabling PS 2.x just to expose more temps? Microsoft said that 12 was enough for PS 2.0 and the HLSL compiler is built around that.

Why not? The R3xx hardware is one of two primary architectures to have support for DX9.

Seriously. How hard could it have been? It's the setting of a DWORD, and the change of a variable in optimization for drivers/compilers.

 

[OpenGL guy]

12 temps is what you get when exposing PS 2.0. If you want to expose more than 12 temps, then you have to go to PS 2.x. What would be the point of enabling PS 2.x just to expose more temps? Microsoft said that 12 was enough for PS 2.0 and the HLSL compiler is built around that.

Why not? The R3xx hardware is one of two primary architectures to have support for DX9.

Seriously. How hard could it have been? It's the setting of a DWORD, and the change of a variable in optimization for drivers/compilers.

From the driver standpoint, not hard. But what about potential application problems? Such as an application that sees "PS 2.x" and thinks "nvidia"?

It's pointless continuing this discussion. Sireric says R300 has 32 temps. I say 32 temps. You say whatever you want.

 

[joe emo]

Why not? The R3xx hardware is one of two primary architectures to have support for DX9.

Seriously. How hard could it have been? It's the setting of a DWORD, and the change of a variable in optimization for drivers/compilers.

OpenGL Guy -- An ATi employee; can probably walk down from his desk and talk to somoene who helped design the r300, or pull up the information right on his computer.


Chalnoth -- Not.


Chalnoth, you do realise your arguing with someone who has infinitely more insider knowledge than you, correct? If I were OpenGL Guy, I'd be higly offended by this. He's already answered your question. I doubt the guys at ATi are like "wow, yeah, Chalnoth on B3D, is right, we should just dop some DWORD edits, and tweak the variable handling in our compiler." I can guarantee the decision for exposing only 12 registers in the r3xx at this time was made carefully and for good reasons.

 

[KimB]

Either way, Sireric's post was just plain wrong.

And btw, it still doesn't invalidate my original point: I don't see how the PS 2.0b profile could possibly be R3xx + F-buffer.

If you'll note the post where I said that, my strongest evidence was the support for only 4 levels of dependency. Using an F-buffer should shatter that limitation.

 

[Bry]

12 temps is what you get when exposing PS 2.0. If you want to expose more than 12 temps, then you have to go to PS 2.x. What would be the point of enabling PS 2.x just to expose more temps? Microsoft said that 12 was enough for PS 2.0 and the HLSL compiler is built around that.

Why not? The R3xx hardware is one of two primary architectures to have support for DX9.

Seriously. How hard could it have been? It's the setting of a DWORD, and the change of a variable in optimization for drivers/compilers.

It made no sense..R3XX still dominated in the DX-9 market even though PS2.0X was released to help NVidia. It could have cause application problems as OPenGL guy said..And with the market leading video card they did not need anything else..Plus we have not seen any games come out that would require ATI to upgrade the drivers to enable the rest of the registers..However now we have SM3.0 out..What a great thing if ATI could release R420 with SM3.0 (although admitingly may have problems but the Checkmark will be there) Still have the best SM2.0 performance (R420 VS NV40) And upgrade the previous generation to SM2.0b so that they can get some of the awsome graphics shown in the Far Cry. In doing this they would be doing two things..Releasing a great card in the R420. and show customers that can't afford to upgrade now that they still are very important to ATI, And upgrading their graphics drivers so they can get some of the goodies too..Talk about customer service..If ATI pulled that off..they would gain alot of customers appreciation that should last for years to come..All I could say then is WOW!!

Although this is just speculation..But I hope it is true.

 

[KimB]

But it still puts ATI in a position of assuming that developers are stupid. I think that this is an asinine position for an IHV to be in.

Oh, and PS 2.x was released not "to help nVidia," it was released so that hardware developers could expose more than base PS 2.0 functionality without going all the way to PS 3.0. This was because Microsoft wanted DX9 to be very future-proof.

 

[OpenGL guy]

Either way, Sireric's post was just plain wrong.

Why don't you just admit your own errors? "If 32 were actually available, they'd be exposed by now." In any event, sireric wasn't really wrong. Have the cap bits changed for R300? No, hence there was no change to the temp registers. There's no practical need to expose four more temps for R300. "Because we can" is not compelling enough to me, and, apparently, to others at ATI.

If the D3D API were slightly different, we would be able to expose all 32 temps, but, again, I don't see how it would help given the PS 2.0 limit of 64 ALU ops. I believe that OpenGL has more flexibility in this regard, especially since there is no 64 instruction limit, but don't quote me on that.

 

[T2k]

Either way, Sireric's post was just plain wrong.

We love you, Chalnoth.