Nv40 16 full pipelines- The Inq.

Any possibliity of a dual GPU design ala Volari?

They had 16 pipes didn't they?


Rampage II!!!

 

That's because there are several things to consider when dealing with a rumor like this. First is the accuracy of your source, but then you have to also consider the accuracy of their sources too.

The Inq is claiming that Nvidia told some person or persons in the US that NV40 had 16 full pipelines and that they then passed said information on to the Inq. Now for this to be true, not only has the Inq, and their source have to accurately pass the information on, but the source has to have received accurate information to begin with. Nvidia's PR has been very slippery about what constitutes a pipeline lately. So we have to consider that even IF the Inq reported the rumor accurately, and that it is what their source was told by Nvidia, does that mean that Nvidia IS using the generally accepted meaning of pipeline or not?

That's why we have to consider Nvidia's credibility as well as the Inq's.

 

NV3x doesn't really either. Ooohh, Ailuros, could you come here for a minute?

Well, I'm still remembering BB asking me "would the enthusiasts love us again if we kicked ATI's ass with the NV40?". I said it would take time. He said I was being too pessimistic. Sounds like he'll be right, and I'll wrong, eh.
Plus, isn't it strange that in late September, they would already be so darn sure they'd kill ATI in this generation? Because, you know, he really was. Seems to me they had expected A2 to be a pretty darn big step-up for a while, eh...

Paul: I guess people can't read
The Inq never ever said "real" pipelines. Let me restate that:

Full pipeline doesn't talk about virtual/real/whatever. It simply means it can output 16 pixels/clock, and that there actually is some processing going on for those 16 pixels, not just "passing" data like the NV3x did with its zixel bypass technology.

And let's not even talk of real/virtual anymore - you guys are just gonna make Ail cry j/k. No, seriously, it doesn't matter. Just call them "real" pipelines I guess, since that's easier.
Also, I speculated that the NV40 was never *meant* to be released at 150M. I would, however, tend to believe many internal NVIDIA documents said it was going to be released in such a design, though, and only the people who had to know it, well, knew it. But that's just speculation, once again.

My point, regarding pipelines, is simple: whatever The Inq currently said is 100% compatible with my earlier suppositions in the other thread. Whether my suppositions are correct, that's another matter. But even if The Inq's article was correct from top to bottom, my thread could also be correct from top to bottom. And thus, for the time being, the only "real" news in that article is the transistor figure.


Uttar

 

You covered so many eventualities in that thread. I'm sure everybody read it with interest, but there was no new information, as such.

Thing is, here you're implying 16x1 under certain conditions without fully specifying them. R300 is 8x1 "in certain conditions".

Your reference to marketing doesn't tally, because when nV said NV30 had "8-pipelines", they were referring to the z-ixel mode. Now we're hearing "16x1" pipelines for NV40 and you're asking us to be cautious (which is fine), but then go on to speculate upon 16 (textured) pixel throughput per clock anyway.



MuFu.

 

Bah, it's midnight MuFu Damn hardware, keeps me awake!
First, saying my NV40 thread just got evantualities means you probably speed-readed it, like others?

The rest was just a HUGE explanation of my reasoning to get to that conclusion... Maybe I should have made that clearer? :?

Read above?

This time around, The Inq used the term "full pipeline". The NV3x has 4 full pipelines, but also 4 "empty" ones. Kinda. If you see it in a physical POV, while it's really virtual (good subject for our next MSN long, boring stupid technical discussion? )

And I've also heard 16x1 from other people too, obviously...


Uttar

 

You can't cover all your bases by using "ifs". That's not how speculation works, mate. You need to say you're 100% confident in something and then if it doesn't turn out to be true you say "see... this is what happens when you talk to loud-mouthed marketing guys instead of loud-mouthed engineers". Then most people go "duuuhh... stupid marketing!".

Believe me, I can't speed read. I need to re-read everything because my brain only retains information for a few seconds before it is replaced with images of naked women and me jumping about in a huge pile of money wearing one of those novelty beer hats.

Corrected, sorry. "For example" is not a specific definition. What I'm getting at is that you can't really make a judgement WRT the "16x1" mode (and possibly write-off the Inq's comments) without knowing what the dependencies are.

Heh, getting ever closer to a satisfactory way to describe CineFX. That's actually not a bad way of thinking about it.

Yah, maybe. I'll do the same thing I always do... smile and nod.

Surely if it's 16x1 in any circumstance then it's a 16-pipeline card by the traditionally accepted definition?

MuFu.

 

with all this talk of NV40 having 16 pixel pipelines, if it turns out that NV40 doesn't even have 8 "real" or "full" pipelines
(as the R300/R350/R360 has 8 pipelines), then that will be a devastating disappointment-- 'cause I see some people on some forums are speculating that NV40 is another 4x2

 

There are various parts of my NV40 speculation thoughts that would tend to allow for a real 16x1 with the quoted transistor figure, but the commentary on the figure including cache in the count might argue against that, depending on how the NV30/NV35 numbers are counted. If counted the same way, the apparent NV31 precedent seems to make it more likely, depending on how the cache sizes compared between NV31 and NV30/NV35 (and how much of the transistor counts are due to caches), how nVidia is doing with their process pioneering this time around, and how vertex processing fits into the design...

The problem with this specific bit of info, however, is the Inqurer's credibility (or lack thereof), which is why I think the idea of this being a leak from nVidia PR has taken such strong root. Not an unreasonable idea at all, by any means , just one not conducive to (relatively) useful speculation. Perhaps the next info leak will provide something useful, or a similar leak from a different source or some bit of info (perhaps already available?) that can provide a new twist.

Wavey's teases are heading in that direction, which has me wondering, but only about the same things already mentioned in the other NV40 speculation thread.

 

....and just wait until next year when speculation starts up about NV50 having 32 pixel pipelines 8)

 

I wouldn't count on that just yet...it ain't over 'til the cards hit retail. 8)

It seems to me that BB ain't really got that great of a grasp on all things technical, he probably just heard some new buzzwords and it got him all excited.

Great, now that mental image is going to be stuck in MY head for the rest of this thread!

 

Oh, and for Wavey's "real" mention, what comes to mind for me is "real numbers", like floating point, versus integer. But even 16 fixed function pixel pipelines would be significant.

 

What about the prospect that NV40 is 16 "unified" pipelines? (ie. each capable of processing pixels or vertices and maybe even capable of switching dynamically). If that is the case then classifying the NV40 as having 16 pipelines (ignoring any Z-only hackery) would be perfectly acceptable to me, even if we would never see 16 pixels/clk in real-world scenarios, since every pipe is made equal. This approach would concurrently satisfy VS / PS 3.0 (and likely beyond).

Such a design is also more probable to fit in a transistor budget of +200M transistors than an NV40 with several vertex pipelines and 16 (slightly simpler) pixel pipes.

I don't know enough about the barriers to implement such a design but I have seen the Microsoft Meltdown slides specifically state that the direction of the Direct3D API is to move towards unifying the pixel and vertex pipeline structure; maybe Nvidia is already there with NV40?

Anyway my hope is it turns out like I'm guessing. I know that ATI was heading in this direction with the original R400, but now we have to wait for XBox2 before that finally comes to fruition.

 

Brian Burke wasn't the man who marketted 3dfx and then when joined nVIDIA started to state how much retarded 3dfx was ?



Nethermind, me too am amzed by Mufu's images.

Tell Humus to do a tech demo..... That would kick some ass !

 

Gah!

 

Could you please all stop with this tiny-speech for a sec? My copypaste buttons are aching...

To Uttar :
I wonder how useful these mutated vertex shaders would end up being in real-world situations... seems to me like there would be a lot of "administrative" performance penalties associated with it ( sort of like how frequent state changes eat up performance in OpenGL )...

 

I think cho might be up to speed.

But then...

?!

MuFu.

 

8x2 isnt even a remote possibility, 6x2 died on the vine.

As far as a 16x1 being Nv40 I say no emphatically. However for the point of speculation if there was a 16x1 card how would it run on AGP or an AGP equivalent bridge? Wouldnt it be bandwidth starved?
Nvidia's not moving to native PCI-Express until the fall remember.

 

j/k

 

THE SIGN!!! ...ON THE BACK OF THE CAR SAID "16 PIPELINES", YOU DUMB ****!!!

Edited 3 times for grammar and spelling.

 

Pipeline guessing aside, if the NV40 truly has over 200 million transistors, bloody hell, that's going to be an expensive chip - especially if it has some high speed GDDR3 stuff stuck on it!

What happens if, when released, the NV40 proves to be say 10-20% faster than R420 in most situations, but ATI is able to undercut the price by 20%?

In other words, if the R420 can't compete in speed, ATI will surely have more leeway to reduce pricing so it is a $400 card as opposed to a $500+ NV40.

That's enough speculation for me tonight.