The Rumor Roundup

[Tim Murray]

Yup, here we go. Sick of misinformation and all that. Best of my knowledge, feel free to correct me. GPU Rumor Watch style.

NV40

NV40 is primarily a 16x1/32x0 architecture.
Reliability: High

NV40 is a 16x1/32x0 architecture at all times.
Reliability: Medium

I don't think it's an 8x2, nor does it seem to abuse vertex shaders. However, there have been a few whispers of conditions that must be met in order for the NV40 to perform as a 16x1 card, so I will include this. I'll look for more information.

NV40 has PS/VS3.0 support.
Reliability: Confirmed, see Beyond3D news

NV40 has gamma-corrected rotated grid multisampling.
Reliability: High for RGMS, medium for GC

No one is entirely sure about the status of gamma-corrected antialiasing, but from all indications we are going to see RGMS for all modes. The fate of supersampling is unknown.

NV40 has angle-dependent anisotropic filtering.
Reliability: Low

From the early screenshots of NV40, it seems that it does, in fact, have ATI-style angle-dependent AF, although not to the extent of R3x0. However, very early drivers were apparently being used, so it is possible that it was simply a driver bug.

NV40 will bundle with Doom 3 or a Doom 3 voucher.
Reliability: High

NV40 will launch April 13th and be available sometime in May.
Reliability: Launch date confirmed, retail availability high (speculation, but almost certainly accurate given past failures with paper launches months in advance)

NV40 Ultra clocks will be ~475Mhz core and ~600Mhz GDDR3.
Reliability: High

From all reports, this is what we've seen on engineering samples, so unless they're gonig to implement a Dustbuster, it should be close to the final clocks.

R420

R420 is eight "extreme" pipes.
Reliability: Was once high, now very low

I hear other things about R420's pipeline configuration. Specifically, how it's not an 8x. The source of that information was an ATI presentation on R420 to AIB manufacturers in November; in the past several months, it seems to have changed significantly.

R420 is a 16x1 card with one quad unit disabled, making it effectively a 12x1.
Reliability: Medium-High

Right now, it's the best guess. It's possible that R420 will be able to output more than 12 pixels per clock in certain situations, but that is still unknown. DOES NOT seem to be vertex shader abuse.

R420 will have FP32 support.
Reliability: High

Crazy, eh? However, no one is sure what will happen to FP24. It could be relegated to _pp, or it could be used for higher speeds on consumer-level cards while slower FP32 is implemented on the FireGL line.

R420 will have memory speeds of ~700-800Mhz GDDR3.
Reliability: High

It's been hinted at for a while. I'm going to go out on a limb and call it high.

R420 will not have PS3.0 support.
Reliability: All but confirmed

Everybody's saying it, everybody seems to know it, so it's extremely reliable. It will support ps_2_x, but it will not support ps_3_0.

R420 will not support VS3.0
Reliability: Speculation

Based on this thread, there seem to be significant problems in coupling PS2.0 and VS3.0, and since R420 doesn't have PS3.0 support, VS3.0 support seems very unlikely. However, R420 supposedly inherited R400's vertex shaders, which were incredibly advanced. Only time will tell.

According to a later post in this thread, the problem with PS2.0/VS3.0 interoperability has been resolved as of DX9.0c SDK Beta 2. R420 will then probably have VS3.0.

R420 will have 8x RGMS support.
Reliability: All but confirmed

It's a given.

ATI will launch second but achieve retail availability first.
Reliability: High

I keep hearing things about GDC, which seem odd, but I do think that ATI will not have anything close to a paper launch. Once they launch, it will be in the store.

I'll update this as necessary. Feel free to tell me I'm an idiot.

 

[epicstruggle]

Anychance of adding rumours on the PCIE counterparts?? Such as whether or not they will be clocked higher?

later,
epic

 

[Evildeus]

That's a good move. thx.
What about the pro/nitro rumor on R420?

R420
- it will give two models of the R420: A pro map with 500Mhz chip and 475Mhz storing act and one?Nitro? map mentioned with 600Mhz chip and 550Mhz storing act.

nv40

- the chip will be feature 32 bit throughout the pipe without loss of speed
- first test with half-life 2 showed that the chip is going to be 2 - 7 times faster than the geforce fx 5950
- the chip will feature RGAA
- the chip will need more power, but will use a single slot solution, as we were told

http://www.nvnews.net/vbulletin/showthread.php?t=26385

 

[KimB]

Feel free to tell me I'm an idiot.

Hey, great. You're an idiot!

Seriously, though, I'm not certain that we can be as sure as you claim that the R420 will not support PS 3.0. However, I will state that the early rumors that the R400 was dropped and the R380 was renamed R420 are evidence in favor of no PS/VS 3.0...

 

[MuFu]

Feel free to tell me I'm an idiot.

Hey, great. You're an idiot!

Seriously, though, I'm not certain that we can be as sure as you claim that the R420 will not support PS 3.0. However, I will state that the early rumors that the R400 was dropped and the R380 was renamed R420 are evidence in favor of no PS/VS 3.0...

The R390 pixel pipeline was based on R3x0. Vertex shader technology was supposedly transplanted from the original R400 project.

 

[volt]

R420 will have FP32 support.
Reliability: High

PS 3.0 requires FP32 doesn't it? It would make sense for ATI not to go with higher precision if it's not supporting it, *unless* ...

Crazy, eh? However, no one is sure what will happen to FP24. It could be relegated to _pp, or it could be used for higher speeds on consumer-level cards while slower FP32 is implemented on the FireGL line.

 

[digitalwanderer]

R420

R420 is eight "extreme" pipes.
Reliability: Was once high, now very low

I hear other things about R420's pipeline configuration. Specifically, how it's not an 8x. The source of that information was an ATI presentation on R420 to AIB manufacturers in November; in the past several months, it seems to have changed significantly.

I'm still hanging with the "12 of 16" pipeline theory, and thinking that 8 "extreme" pipes means that it can do an 8x1 32bit mode.

R420 will have 8x RGMS support.
Reliability: All but confirmed

It's a given.

What is "RGMS"?

ATI will launch second but achieve retail availability first.
Reliability: High

I keep hearing things about GDC, which seem odd, but I do think that ATI will not have anything close to a paper launch. Once they launch, it will be in the store.

I'll update this as necessary. Feel free to tell me I'm an idiot.

You're an idiot, but I think you're close to right.

BTW-Hanners is gonna kill you, he's been working on a Rumor Round-Up piece all afternoon in our private staff forums!

 

[KimB]

- first test with half-life 2 showed that the chip is going to be 2 - 7 times faster than the geforce fx 5950

That's, um, a big range of possible performance.

 

[digitalwanderer]

What is "RGMS"?

Rotated grid multi-sample anti-aliasing, never mind....I'm having a weird day and just getting back to the viddy card thang.

 

[volt]

Diggers,

RGMS - Rotated Grid Multi Sampling
OGMS - Ordered Grid Multi Sampling

 

[KimB]

R420 will have FP32 support.
Reliability: High

PS 3.0 requires FP32 doesn't it? It would make sense for ATI not to go with higher precision if it's not supporting it, *unless* ...

I don't think there is any difference in the requirements.

Oh, and Baron:

Crazy, eh? However, no one is sure what will happen to FP24. It could be relegated to _pp, or it could be used for higher speeds on consumer-level cards while slower FP32 is implemented on the FireGL line.

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. As for FP24 support in the future, I think ATI has had enough of a history by now of using older technology for low-end parts, so we'll see the R3xx continue to be used, and continue to shift more towards the low-end of the market.

 

[KimB]

One minor addition:
RGMS isn't an accurate description of modes other than 2x and 4x.

 

[MuFu]

I'm still hanging with the "12 of 16" pipeline theory, and thinking that 8 "extreme" pipes means that it can do an 8x1 32bit mode.

Have you heard that from anybody else apart from me in the other thread? I don't think it's very likely.

 

[Joe DeFuria]

The only significant quibbble I have is the speculative memory speeds.

We're thinking that NV40 will be primarily 16x1, paried with 600 Mhz Ram, whereas R420 would be 12x1...with 800 Mhz? And given the history of ATI not shipping with bleeding edge ram, makes me personally think that 800 Mhz Ram is extremely low on the probability scale.

I personally think your NV ram numbers are pretty good...approx 600 Mhz for the ultra, but I'm thinking ATI will be closer to 500-550.

EDIT: Also, I'm on the fence about the move up to 32fp in the pixel shaders. If it is 32fp, I'm guesing 24fp would be dropped altogther.

 

[volt]

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

 

[Joe DeFuria]

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?

 

[digitalwanderer]

I'm still hanging with the "12 of 16" pipeline theory, and thinking that 8 "extreme" pipes means that it can do an 8x1 32bit mode.

Have you heard that from anybody else apart from me in the other thread? I don't think it's very likely.

The "12 of 16" bit, yes; the "extreme" meaning 8x1 32bit, no...I just like it and it's the only way I can make the whole bloody 8 "extreme" pipes fit into me "12 of 16" pet theory and the 8 "extreme" pipes has just been coming at me from too many reliable sources for me to discount.

Can you think of another way to dovetail the two theories together? :|

 

[Tim Murray]

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.)

Chal: I agree, FP24 as _pp doesn't make a whole lot of sense. However, if MuFu is correct in thinking that R420's PS units are FP24 or FP32, with 12 FP24 ALUs and 8 FP32 ALUs, there would be a performance increase from using FP24 as _pp. However, I tend to think that's nuts and even if they are both FP24 and FP32, one will be selected in the BIOS with no _pp available.

Evil: Only place that's mentioned a R420 Pro/Nitro/Whatever. I tend to believe that there will be two versions, just like with 9700 and 9800, and they will be Pro and non-Pro. Of course, they might have a card with extremely high core clocks and pair it with a two slot cooler.

Epic: That's too far away for me to predict either right now. I'd say that in the case of PEG NV40 (or NV45 or whatever you want to call it), it's probable that it will receive at least a memory speed bump. With R420/R423, I'm not sure; maybe the four extra pipelines? However, that'd be a bit nuts, seeing as how you'd definitely need a bigger cooler to maintain the same clock speeds. Perhaps R423 will also include R360-like improvements, basically tweaking the process and reducing thermal output somewhat.

 

[991060]

About the VS3.0/PS2.0 pairing problem，I think MS still loves ATi. According to the latest DX9.0c SDK beta 2 release, the problem has been resolved. On the other hand, this fact does enhance the credibility of R420's not having PS3.0 support, coz if it's not, making VS3.0 and PS2.0 compatible is not very necessary.

 

[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).

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.