[AnarchX]

G92:
http://www.vr-zone.com/articles/Nvid...V770/5645.html
G94+G96:
http://www.vr-zone.com/articles/Nvid...iled/5705.html

Hope, we will see this GPUs on desktop, too.

[Ailuros]

Uhhhm that table doesn't show anything considering G92b or I'm missing something.

[AnarchX]

You are right, but VR-Zone also reported earlier about 55nm G92b.

[Ailuros]

Frankly for desktop SKUs it'll be only really interesting if the b/55nm parts come with higher frequencies than those on 65nm.

[AnarchX]

Nvidia First 55nm Desktop Graphics; GeForce 9800 GT

With rumored 650/1625/1000MHz it should be still a lot cheaper to make for NV, since smaller die and hopefully lower consumption.

[Blazkowicz]

Originally Posted by Ailuros

Frankly for desktop SKUs it'll be only really interesting if the b/55nm parts come with higher frequencies than those on 65nm.

I always like lower power.

[AnarchX]

G96B = 121mm²

G84 	-> G96A 	-> G96B
80nm	-> 65nm		-> 55nm
169mm²	-> 144mm²	-> 121mm²
100% 	-> 85%		-> 72%

What a bad shrink?

[Domell]

So GF9800GT and GF9900 will be based on G92B?

[w0mbat]

Hm, now nV is really going to mess up everything even more -.-

8800GTS, G80, 90nm
8800GTS, G92, 65nm
8800GTX, G80, 90nm
9800GTX, G92, 65nm
9800GT(S), G92b, 55nm
GeForce 9900 = GT200?

[INKster]

Originally Posted by w0mbat

Hm, no nV is really going o mess up everything even more -.-

8800GTS, G80, 90nm
8800GTS, G92, 65nm
8800GTX, G80, 90nm
9800GTX, G92, 65nm
9800GT(S), G92b, 55nm
GeForce 9900 = GT200?

I still have strong doubts that the "9900" moniker will ever be used by Nvidia (at least outside the mobile space).

[Arun]

Well keep in mind that's within the margin of error of the expected 19% shrink. 148mm² for G96A and 120mm² for G96B would be ~exactly that. However, what's most stunning is die size relative to G84. We're talking about a ~30% shrink for a a full process node shrink! That's not just bad; that's godawful. Talk about underdelivering, ugh...

It's fair to say NVIDIA screwed up the 65/55nm transition as badly as ATI screwed up the 90/80nm transition, if not worse. In a way, it could be said they also screwed up transitioning from G80 to "absolutely anything else". Both G84 and G86 were pretty good, but going from ~480mm² on 90nm to ~170mm² on 80nm for 1/4th the shader performance and 1/3rd the ROPs/MCs is subpar even if triangle setup etc. remained identical/similar.

Either way, G96B competes nicely with RV635 given they both have the same die size on the same process node. The only problem, of course, is that once RV730/RV740 are out there, NVIDIA might just be in a lot of trouble. They sure as hell better start praying those chips won't be out there in time for the Winter OEM cycle. As for RV710, at least they do have a viable competitor to that in the wings...

[mczak]

I still don't get the G84->G96A->G96B transition. Does G96A even exist? Apparently there were sample 9500GT boards using it, but I've never seen one you could buy. Or does it exist as a mobile gpu only? What a weird strategy. Or given the less than stellar 80nm->65nm shrink, maybe G84 (using old tech) is cheaper to produce so nvidia just kept going with that?

[Arun]

G96A had issues AFAIK, at least at UMC which presumably was the sole foundry. I don't know what foundry G96B is at, presumably TSMC but certainly it might help explaining the subpar density if it's UMC-only although I don't believe that it is. There are also leaks that implied the chip exists at both foundries. Conclusion: Nobody knows what the hell happened with both G96/G98. It's fair calling both of them complete and utter fiascos though.

You know, I wonder what the gate length for both RV770/G92b/G96b are. I think the subpar clock scaling on G9x would tend to point at using a lower-leakage kind of transistor for G9x, which might also have higher gate lengths/lower density - Hmm! Either way, that's one decision that certainly doesn't look like it paid off if so, to say the least...

[AnarchX]

And G94B:
http://en.expreview.com/2008/07/28/n...ging-the-name/

Still a big die in comparison to performance.
Lets see how small RV730, the natural competitor, will be (atm I estimate something around 150mm²).

[trinibwoy]

Wow...why did they even bother?

[Arun]

Seems like a ~17.5% shrink to me, about what you'd expect I guess. Compared to RV670, it's a small die, but unfortunately for NV that won't be the point for very long! (although if they got the back-to-school OEM design wins based on that, it'd be positive for margins at least) - as for why they'd bother, why not? Remember mask costs are just a few million dollars. In the grand scheme of things, it's literally cents. If their R&D expenses for doing this were minimal, they'd be madmen for not doing so (as to whether that's the case, who knows!)

As for RV730 being 150mm²... uhh, what bus width are you expecting it to be? Because if it's 256-bit, that'd be massively pad limited.

[AnarchX]

Originally Posted by Arun

As for RV730 being 150mm²... uhh, what bus width are you expecting it to be? Because if it's 256-bit, that'd be massively pad limited.

There are reports about 128-bit, what seems likely and should be enough, if we look on HD3690. Maybe later RV730 will be combined with GDDR5.

[mczak]

Originally Posted by AnarchX

There are reports about 128-bit, what seems likely and should be enough, if we look on HD3690. Maybe later RV730 will be combined with GDDR5.

Isn't the pad limitation more a result of outline, rather than area? So you could gain more pads by making it non-square (for large chips, I'd gather this would be impractical, but for smaller ones could be just fine?). Not sure if for the RV730 timeframe cost would allow to use gddr5 on such a cheap part (well it's hard to say since I haven't seen any quotes what gddr5 costs in comparison to gddr3 and how prices are expected to move relatively).

[trinibwoy]

Originally Posted by Arun

Seems like a ~17.5% shrink to me, about what you'd expect I guess. Compared to RV670, it's a small die, but unfortunately for NV that won't be the point for very long!

I guess. It will be interesting to see how clocks/power consumption turn out. What's the current expectation for RV730? 4 SIMDs? G94b would probably have trouble keeping up but that's about the best Nvidia can muster for a while it seems.

In general Nvidia seems to be flubbing all over the place. First they make a lot of noise with G92b and 9800GTX+ and now 9600GT is getting a stealth shrink to 55nm. They really aren't inspiring confidence at the moment with all this confusion and inconsistency.

[mczak]

Originally Posted by trinibwoy

I guess. It will be interesting to see how clocks/power consumption turn out. What's the current expectation for RV730? 4 SIMDs? G94b would probably have trouble keeping up but that's about the best Nvidia can muster for a while it seems.

Why do you think G94b couldn't keep up with rv730? I wouldn't expect the latter (in a 4 simd configuration) to be any faster than rv670 (at similar clocks) - in fact it should be much slower in some areas (like the half-rate fp16 texture filtering). And G94 seems to compete quite well with the HD3850 already. If rv730 only has a (gddr3) 128bit memory interface, I'd suspect it's almost certainly going to lose against G94b (in 256bit configurations) (though in this case it should also be cheaper).

[Jawed]

Don't rumours indicate there'll be RV730 and RV740? If so, 128-bit and 256-bit?

If RV740 is a cut-down RV770, e.g. 6 clusters and no CrossFireX port, it should be about the same die size as G94b I guess.

we might see RV740, if it is indeed this spec, being very close in performance to HD4850. Sort of a re-run of the way 9600GT ate into 8800GT.

Jawed

[trinibwoy]

Originally Posted by mczak

Why do you think G94b couldn't keep up with rv730? I wouldn't expect the latter (in a 4 simd configuration) to be any faster than rv670 (at similar clocks) - in fact it should be much slower in some areas (like the half-rate fp16 texture filtering).

Hmmm you may be right. I underestimated how well G94 does versus RV670 today. Just looked at some recent benchmarks and it competes quite well (I blame Nvidia's mess of a lineup for my lowered opinion of the 9600GT).

[Speccy]

Probably DisplayPort, for one.

[CarstenS]

Some improvements in ressource utilization, better scheduling, larger register file (even more crucial IMO the less ALUs you have), maybe even recovery of the lost MUL?

[Putas]

Most likely, since it is just hair slower then 20% higher clocked 8600 GTS.