Nv40 16 full pipelines- The Inq.

All IMO and perhaps a bit speculative, but:

A) The NV40 won't be capable of loopback (or at least, probably not, and if it's the case, only in very specific and rare cases) under 12x1 and 16x1 mode. That's similar to the NV31/NV34/NV36 "double-pumping" technology.
B) Keep in mind the number of ROPs remain key. Under 8x MSAA, unless it has 128 ROPs, which is rather unlikely IMO , it won't be capable of 16x1 operation, and probably not even of 8x2 operation. More like 4x4. But that's also the case of the R3xx and R4xx: Dave's R300 MSAA thread revealed us a few months ago that under 4x MSAA, it was *practically* operating in 4x2 mode.
C) Considering the two above points, if the R420 is 8x2, and the NV40 is 6x2/8x2/12x1/16x1, the NV40 would in fact currently have the least impressive spec, as it wouldn't have access to 16 texture lookup units non-stop, while the R420 does...

Seriously, this pipeline crap is not only useless, it's also boring. Now, if we were talking about arithmetic performance and microarchitecture.......

Ailuros: Hehe, you love taunting them with your heat dissipation figures, don't you?


Uttar

 

Yes, but nV40 would have to be an all new fresh design.

I mean, if they didn't go the pure new architecture way, the nV40 will be in the same direction as nV3x in transistor consumption......

 

Well we discuss rumours available to us... If you can throw some new ones, I think everyone will be glad to shift to more complex rumourish-speculations

 

R360 has 107m transistors as compared to NV38's 130m. This means that NV38 has around 21% more transistors yet we know performance isn't better in many situations, especially with DX9 shaders.

If the rumours are correct, then R420 will have 175m transistors and NV40 will have 210m, once again a 20% advantage for the NV chip. Therefore we shouldn't assume NV40 will be faster.

Of course, this all assumes that the 130m transistors in NV3* were all functional! If large chunks of the chip didn't work then a proportionally larger increase in performance could be expected with NV40 - as long as all parts of this chip work!

 

Why should they have kept non-functional blocks in the NV35? So imho your argument is rather pointless.

 

There is no doubt that 16 real pipelines are possible with 200+ million transistors heck 16 real static pipelines would be possible with 80 million transistors.

The real question is not how many pipelines, but what are the pipelines capable of.

 

My argument is about the (rumoured) proportional difference in size in transistors for either chip. Just saying that more transistors does not necessary = increased performance/functionality.

The blurb about 'non-functional' blocks was just an aside which took into account that some people are claiming the surprisingly average performance of NV35/38 is down to it not working properly. Personally, I don't believe that this is true or, as you say, they would surely have fixed it for NV35/38.

All speculation anyway.

 

edit: naah

 

Because Huang had already stated that work on NV3x had all but ceased (bar minor tweaks) - nVidia were focussing on the next generation of chips ie: NV40 instead.

 

DC, good point: ATi packed 8x1 into 107M. But for nV to do the same would require them to basically change architectural philosophies and realities, no? FP32+16 and separate FX12 into just FP24? Shaders over textures by going with one "TMU" per "pipe"? I guess I find it hard to believe they'd do so. I suppose they changed after NV1, so they may do so again. But GPUs seem to be getting awfully complicated to pull off a switch like that without huge risk. Can they, at this point?

Returning to the subject of transistor counts, wouldn't the extra logic PS/VS3.0 entail require more transistors above the R300's supposedly bare-minimum DX9 2.0 figure? I mean, assume we double 107M to achieve essentially a 16x1 DX9 PS/VS2.0 card. Can they fit the extra logic required by PS/VS3.0 into that space, even assuming they ditch separate FX units and perhaps remove a VS unit or two?

I'm still curious how nV will be able to compete with ATi in speed/heat without low-K, if what ATi says is true and they were able to achieve 500MHz on RV360 at the same energy cost as 400MHz on RV350.

(Now, what was that I said about the sidelines? :\)

 

Edit:
No need for my post anymore, nothing to see here, please disperse...

 

You're not willing to bet that it will go unnoticed and won't cause a huge load of debates and practical jokes after release, are you?

Please....you're just making my headache with all the number crap worse than it already is. Take a somewhat 3*NV35 scenario, it's simple enough, people can understand it and we'll see afterwards where weaknesses and strenghts are.

By the way if clock/ram would be hypothetically the same on R420, I don't see why tri-TMUs wouldn't make sense vs a 2 TMU per SIMD scenario.

We're not going to get behind the real specs of either/or until announcement anyway and we're all going to be wrong one way or the other. Let me shake the foundations of your NV40 X*Y theories a bit: expansion based on RezN8.....you were saying? 8)

 

Just had a crazy idea.

What if dave meant 16 real pipes as physical pipes i.e heat pipes.
This thing might be an external card with its own case and power supply!


[EDIT] J/K

 

anaqer, SFF's are a niche, no matter how you spin it.

My TT isn't really that loud. The case is well shielded, and the fans are throttleable via knobs on the front. I would be annoyed if my X-Box or PS2 was as loud as the average PC, but those sit on a shelf in the living room. My digital projector is far louder than any dustblower you could ever conceive, but then again, it's got a bulb that outputs 2200 lumens.

I just don't buy the argument, and I didn't buy it much when the NV30 came out. Anyone concerned enough to pay $400-500 for a video card is not buying it because it looks nice and sounds quiet. They are buying it because it's fast.

It's bad enough that the market for $400-500 cards is a niche market, but now we're contemplating people shoving Prescott's and super-high-end GPUs into a SFF box, and expecting it to be cool and quiet? This bears a resemblance to the HD-editing thread we had. People making arguments based on very rare, but so-called "revolutionary" use cases for GPUs.

If the NV41 is loud and hot, I think people have a point. But I feel the highest end GPUs are concept cars, meant to demonstrate pure performance, to take the crowd. They're Formula 1 cars.

 

what do you think will be the clock speed diff from a 160 mil+ Low-K 0.13 to a 210 mil plain 0.13?

Edit: asuming similar design teams...

 

Yes they are - right now. I suggest we get back to this in a year or two ( maybe check some of the Dell models...? )

$500 cards are like luxury items - you expect them to GIVE you something for the price. Note that I said "for the price", not for the noise - you have already sacrificied something ( ie your greenbacks ) for the performance, you don't want to gulp down the noise levels of a screaming banshee too.

Nobody's expecting a Prescott OR a super-high-end GPU in a SFF box. You seem to be missing the point: the average user does not need OR want that. I know it sounds stupid to power users and geeks, but they tend to forget what a minority they actually are. You know what I'm seeing for the next-gen SFF boxen? Desktop Banias with RV410 and NV41. Performance freaks need not apply - they are NOT where the big money is.

 

Ain't that the truth! This is the most sensible comment I've seen in these forums for a long time.

 

Depends on what exactly those extra transistors were spent on, whether the "plainness" is for real, whether they decide to do some factory overvolting, etc. - my guess would be ~450MHz ( don't know what the original target clock was, 500MHz or 600MHz ?).

 

I just said that in the post you responded to.

The noise, heat, and power matter for the mid-range parts. The high-range parts are meant to secure the performance crown and satisfy hardcore users, but they aren't the bread and butter.

If someone can come up with a dual-PCI-E SLI solution wherein I can plugin 2 or more R420s or NV40s and link them together, I'm there, heat, power consumption, and noise be damned.

 

Well, of course not! Those figures are absolutely ridiculous indeed, and so far, do seem to be the biggest problem NVIDIA will have to deal with. I pity whoever's in charge of designing their cooling solution, really...

But even more catastrophic IMO is the NV41 figure. I mean, come on. They might as well completely give up on the notebook market now, because if they want to get reasonable fifth tenth of their desktop clockspeed?

Err, WTF? It's more like 1.75x or maybe 2x NV35 on a per-clock basis... 3x is a *massive* exageration. It's not 12x2, you know Although, if we're talking about arithmetic performance, that's another beast completely. But we don't know enough about it yet.

What? What's RezN8? :s Sorry, must be some type of thing Americans or British people know about... But I don't. And yes, I'm an idiot
Seriously, I know those are not real pipelines. There would be two PS pipelines and one VS pipeline. 3x2+3x2+2x2, if people still want to speak in those terms.

I don't know how many times I've complained about how sucky X*Y is, and if I'm still using it, it's because it's the only information most sources understand (*cough* The Inq *cough*). So getting anything else is extremely hard.

Now, don't get me wrong, I'd love to be able to speak in REAL architectural terms. We probably got enough info to do so for the NV3x now, and we can assume it's roughly similar for the NV4x (although I wonder how much better their gatekeeper and/or "scheduler" have gotten). But if I talked that way, I'd be happy if 2-3 people would understand me, so...

But yes, you're right, we don't know the real useful details. Which is why, for now, we probably should just shut up. Eh.


Uttar