NV41 - Geforce 6700

[skazz]

Can anyone give me a subtle hint as to when the NV41 cards are likely to show up in reviews?

Now I try not to take anything they say seriously, but I couldn't help noticing that the InQ (http://www.theinquirer.net/?article=19179) seems convinced that NV41 a.k.a. 6700 will start rolling off the production lines in November. Given the paper launch approach of recent months, that means we should be seeing reviews of "almost final silicon" of the 6700 in the next week or so?

 

[Diplo]

Presumably, if it really exists, the 6700 will sit somewhere between the 6600GT and 6800. Which begs the question, is there enough room?

 

[Kaotik]

The clocks Inq posted can't be true, with 128bit bus 12x1 chip at those speed would be too bottlenecked, with 256bit it would be kicking 6800NU's arse all over the place, which wouldn't fit the numbering scheme at all.

 

[Pete]

It's not that much faster than a 6600GT on the core. How much of a bottleneck will 1GHz RAM be with shader-heavier gamers?

It sounds like the 6700 would make for a nice replacement for the 6800, so that nV could concentrate on shoveling NV40 cores to 6800GT cards.

 

[Lezmaka]

The clocks Inq posted can't be true, with 128bit bus 12x1 chip at those speed would be too bottlenecked, with 256bit it would be kicking 6800NU's arse all over the place, which wouldn't fit the numbering scheme at all.

And that'd probably be because it's supposed to replace the 6800NU. As yields on NV40 get better, that would seem to also mean more fully functional cores would be used to fulfill 6800NU demand. So, you make a new core that only has 12 pipes and that also means you can get more chips, but still sell them at a similar price and get more profit. That would also leave all the fully functional NV40 cores to be sold at GT/Ultra prices instead of NU prices and getting more profit.

 

[skazz]

Sorry... didn't bother with additional comments...

NV40 models 6800NU/6800LE are indeed no longer needed once the NV41 appears, since if you believe Inq it will be basically a slightly cut down NV40 core (i.e. 12x1 / 5 VP) but slightly faster core speed than the 6800NU currently has (and higher RAM speeds). The 128MB variant of the 6700 would outperform the 6800NU by a small amount, but even the 256MB 6700 variant would not reach 6800GT performance. Both 6700 variants would outperform the 6600 series by a considerable margin, of course.

(Comparing to the 6800GT, a 256MB 400/1000 12x1 6700"GT" would have identical RAM bandwidth but wouldn't reach the same kind of fill rate, since 12x400 < 16x350)

However, I think the answer is very simple : 6800NU/LE will remain AGP until stocks dwindle, whereas 6700 128/256MB variants will start on PCI-express and may never arrive as an AGP variant, giving NVidia leeway to continue with the 6800NU/LE as long as is necessary depending on their NV40 yields. If you believe the Tom's Hardware article of a few days ago NVidia and ATI are both suffering 30-40% yield on their top end products, so NVidia will continue to need the 6800NU/LE option for some time to come.

I was just thinking that a 6700GT with SLi would make a very nice addition to a new NForce 4 SLi mobo, particularly if it turned up at around the 325-350 euro price range

 

[deadly sk1llz]

In addition, I believe it has been mentioned somewhere that the price point of the 6800NU and 6700 will be the same. (Around $250-300 US) It's a pretty good absumption that 6700 with be only PCIe and 6800NU will be AGP still.

One point made is that more and more 6800NU's are being flashed successfully to a 6800GT (newer 6800nu boards that is), so they feel that nvidia is losing profit by not just making the perfectly fine 6800 boards into a GT or Ultra model, but having to downgrade it to a 6800nu, just to have them in stock. I don't know if this is true, or if this is even a valid point, but it is an interesting coincidence.

 

[Pete]

1GHz mem leads me to believe it's 128-bit. Is it not?

 

[ChrisRay]

1GHz mem leads me to believe it's 128-bit. Is it not?

Why? I dont doubt that reducing the memory from 128 from 256 will radically reduce cost below the GT, though I bet we end up seeing something more conservative like 900 Mhz or 850.

 

[Pete]

Wait, you do or you don't doubt?

500MHz 256-bit DDR seems like a lot for a 400MHz, 12-pipe core (=4800MP/s)--at least, compared to a 350MHz, 16-pipe (=5600MP/s) core. It's a huge step up in memory bandwidth for a relatively small step up in fillrate, too, compared to the 6600GT (500MHz * 8-pipe = 4000MP/s).

If it is 256-bit, are we looking at no more 6800 and 6800GT, considering how close a 400MHz, 256-bit 6700 will be in terms of performance? Will even the 6800U be cancelled?

Or maybe NV41 can clock higher...?

 

[ChrisRay]

Wait, you do or you don't doubt?

500MHz 256-bit DDR seems like a lot for a 400MHz, 12-pipe core (=4800MP/s)--at least, compared to a 350MHz, 16-pipe (=5600MP/s) core. It's a huge step up in memory bandwidth for a relatively small step up in fillrate, too, compared to the 6600GT (500MHz * 8-pipe = 4000MP/s).

If it is 256-bit, are we looking at no more 6800 and 6800GT, considering how close a 400MHz, 256-bit 6700 will be in terms of performance? Will even the 6800U be cancelled?

Or maybe NV41 can clock higher...?

I dont think the 6800 GT would have anything to worry about in regards to a 300 dollar priced 6700 (Assuming thats where this is the target price.)


6800 GT

5600 Fillrate

32 Gigs of Bandwith,


6700

4800 FIllrate,

32 Gigs of Bandwith (Up in Air)


6600GT

(4000 Fillrate)

16.6 gigs of bandwith.


I mean the 6800NU by default actually has 3900 in fillrate which is less than the 6600GT assuming you dont take into account the pipeline differences and only being able to do 4x pixels output per clock I think its entirely possible to have the 1 Ghz Memory, In the long run its still probably cheaper than selling 6800NU's with an NV4x core with cheap memory. So to answer your question. Both, I believe its entirely possible but they could opt for something in the middle (what I'd do )


Either way this is a nice replacement to the 300 price point and likely to put up some impressive numbers. Now, the 6800NU is probably one of the worse values atm. It's just a bit over the old generation, And usually quite a bit under the new generation (large discrepency in Fillrate for example) So I find it very likely they'll replace it with something more in lines with performance differences.

 

[skazz]

All very true, but since over 50% of all 6800NUs (taking the hundreds of votes on various forums as roughly realistic sample) seem to be softmoddable to 16 pipelines, that lack of fill rate has become less of an issue for many 6800NU (and 6800LE ) owners.

Looking again, without worrying about overclock abilities:

6700
4800 Fillrate
16,6 of 33,2 GB/s bandwidth (depending on 128-bit or 256-bit)

6800NU non soft-modded
3900 Fillrate
22,4 GB/s bandwidth

6800NU soft-modded
5200 Fillrate
22,4 GB/s bandwidth

Now let's add the results of an easy overclock for the average 6800NU (360/850):
5760 Fillrate
27,2 GB/s bandwidth

I'd say there is no way 6700 will come out with 128-bit memory bus, since it would impose far too great a bandwidth restriction. Even with huge RAM overclock there is no way the 6700 would get even close to the memory bandwidth capabilities of the 6800, and I can't believe NVidia would do that.

At default settings, the 6700 is a clear step ahead of the 6800, definitely, but both will suffer the same 128MB RAM bottleneck which will be the main reason for falling behind the 6800GT. A softmodded 6800NU gets back in front of the 6700 in terms of fill rate but still suffers from lower bandwidth compared to a 256-bit 6700. I therefore agree that NVidia might well end up with at least one 6700 model with slower RAM. Of course, maybe NVidia has too much 1 GHz DDR-3 RAM at the moment, and can therefore include it without problems on the top 6700 model despite it not really being necessary?

How about this kind of approach?

6700GT = 256MB, 400/1000
6700 = 128MB, 400/800-1000
6700LE = 128MB, 350/700, 8 pipelines, OEM only (direct competitor with the 6600GT of course)

 

[Dave Baumann]

I would wager that there would be one specification that matches the current 6800 speeds. I think that there is a possabuility that 6700 will knock out the 6800 (standard) SKU, but then if you have a 6700 SKU that matches the current 6800 (standard) spec you can freely interchange the chips between NV41 and NV40 fall-outs.

 

[skazz]

Sounds extremely sound business approach... however I was wondering if there is also an ATI Christmas refresh product which NVidia will need to challenge at the 300ish Euro price point?

The 6700 may come out with higher clocks than the 6800NU at 300 euros if there is an ATI card needing beating, of course.

(edit: e.g. the R430 16 pipeline competitor? http://www.hkepc.com/bbs/viewthread.php?tid=229551 )

 

[Pete]

I would wager that there would be one specification that matches the current 6800 speeds.

Bandwidth included? I'm still doubting a 256-bit bus on this card.

Also, any word on ROPs? We lookin' at eight, or a more budget-conscious (albeit possibly too limiting) four?

You can tell I'm strongly considering this a more value-oriented card, but I guess if ATi and nV have sold 256-bit bus cards at $200-300 for the past two years, they'll probably continue to do so, even with 1GHz GDDR3.

 

[Dave Baumann]

If this is a 130nm part then I would say this was be 3 quads all the way through and the die size would probably sufficient for 256-bit.

 

[Ailuros]

I have the weird feeling that success rates with softmoding on future NV41 will shrink by quite a bit.

Under the presupposition that there will be "two families" of NV41's; one being NV4x's with disabled units and one being genuine 3 quad boards.

In that case no a 128bit bus does not make any sense and NV41 is for the record considered by NVIDIA part of the high end segment last time I checked.

***edit: ROPs are one of those strange riddles for future NV41s . It's my understanding that on 6800nonU's despite the disabled quad, all 16 ROPs are still operational.

 

[Pete]

Why do I think NV40 has 8 ROPs? Logically, it should have 16, if only because it needs to in order to be 32x0. Am I confusing alpha blends with pixel outputs? Where are alpha blends done in the pipeline, anyway?

Edit: Right, right, it has eight color ROPs and an additional eight Z ROPs. Interesting that nV seemingly simplified things with the NV43 and now the NV41; more wiggle room in the transistor/die size budgets?

Anyway, it sounds promising as a cheaper 6800, but I'm guessing nV is making it to boost their profits at that price point, rather than to lower the 6800 price point. It'll pretty much have to debut at MSRP$300 or above, as the 6600GT appears to be MSRP$250 (according to eVGA's site, anyway, which may be overstating things, as I don't recall nV cards selling for $50 below list right off the bat).

If a 6700 at 6800 speeds does slot in at $250, b/w a $200 6600GT and a $300 6800, then it should sell very well in the AGP arena. But then, this'll probably be PCIe first/only...?

 

[Xmas]

Why do I think NV40 has 8 ROPs? Logically, it should have 16, if only because it needs to in order to be 32x0. Am I confusing alpha blends with pixel outputs? Where are alpha blends done in the pipeline, anyway?

Alpha blends are done in the ROPs, but NV40 can only blend 8 pixels/clock.

Just as a side note, NV40 can not do 32 Z-compares per clock. These have to be done in the earlyZ stage.

 

[Pete]

Sorry to further pester you, Xmas, but I'd just like to be sure:

1) Alpha blends are limited by full, or "C"ROPs, right?
2) But the 32x0 figure still stands, as with NV30's 8x0? You're just further parsing the umbrella term "z-ops" rather than directly contradicting 32x0, right? Or are you saying NV40 isn't as completely 32x0 as NV30 is 8x0?

TIA.