[Headstone]

So according to the above Cascading Crossfire plan there will be only 2 PCBs each with 2 gpus onboard. Now what are the chances that Asus will be releasing that dual x1600, or is it just a POC? Now what would the cooling requirements be for a dual gpu dual card crossfire setup be, let alone the power requirements...

Now they say that Canada is cold but a setup for the x1900xtx like that (or at least the x1800xt) could heat up the country more than the current leadership campaign

[overclocked]

Originally Posted by Chalnoth

Bandwidth and latency. The longer a bus has to be, the harder it is to drive it at high frequency. It'd also be rather expensive to have a 256-bit interconnect between two cards. And then there's the need for many additional pins on the chip's packaging. And the latency would be greater no matter what you'd do.

All that said, maybe one could make use of the two video cards in such a way that one uses the other for sort of a cache that sits between the card and the system memory. But it'd be rather difficult to find good, efficient algorithms for sharing data between the two cards in such a way that data is stored in local memory in the vast majority of cases where it's needed. And even then, bear in mind that the connection between the two cards is most likely of lower bandwidth than the connection between each card and system memory, so it might be better to just not bother.

Thanks for the good reply Chalnot.
In other words this isnt something that ever will be feasible IYO in a good way because of the technical hurdles?
Its quite interesting as specially on one board with two gpus it seems like such a waste.
You dont think some kind of tiling algorithm could be used?

[Chalnoth]

Yes. And now that I think about it, I think that the primary problem with sharing memory would be in the interchip bus. Basically, it's extremely expensive to make that bus much larger than what is absolutely necessary to composite the frames together. So even if you could find a really efficient algorithm for splitting texture data between the cards, there would probably be more bandwidth available for transferring textures from system memory than there would be between the cards.

[Redeemer]

How can Nvidia make 32 pipes @ 700-750 available? ATI's low pipeline high shader throughput design @ reasonable clocks makes all the more sense now after their R580 launch. Is there anything else Nvidia can do besides adding pipes?. Obviously what good is a product with no availability. Of course G71 specs are not in stone yet or is it?

[jb]

I am sure that the G71 will be a very excellent card. We still have not head many concrete rumors on its specs. But those that are in the know say things will be interesting this year. So take that as you will...

[Carl B]

What do you mean by reasonable clocks though? No more or less reasonable in a sense than what ATI has reached with it's x1800 and x1900 parts, and they still have the drop to 90nm (80nm?) on the high-end to work with. But anyway, I would certainly understand if they fell short on the ~700 target. That being said, I don't think the 32-pipes should present any more or less of a yield issue than it would for R580. Going down to 90nm, even with the addition of up to 100 million transistors, die sizes should in fact improve.

Anyway I think we're getting a little ahead of ourselves with G71 yield predictions.

[digitalwanderer]

I'm pretty sure the G71 availability is extremely non-existant.

[neliz]

Originally Posted by digitalwanderer

I'm pretty sure the G71 availability is extremely non-existant.

yeah.. it's pretty hard to find one right now..

[no-X]

I'm not sure if this combination (32pp + 750MHz + 90nm) describes G71 accurately, but the more colossal specifications, the worse availability. GTX->GTX512 performance boost (expecially with FSAA enabled) was caused mainly by more than 40% faster memory. I doubt, that nVidia will use anything faster than 1.1ns GDDR3 modules. And increasing texturing power without adequate bandwidth could lead to lower-than-expected performance. I can be wrong, of course, and nVidia could be waiting for speedier GDDR4. In this case, the performace could be brilliant, but it's availability and price not.

[digitalwanderer]

Originally Posted by no-X

nVidia could be waiting for speedier GDDR4. In this case, the performace could be brilliant, but it's availability and price not.

The G71 is going to have to be widely available, anything less on their launch day and it will be viewed as a spectacular failure.

[Ailuros]

Originally Posted by jb

I am sure that the G71 will be a very excellent card. We still have not head many concrete rumors on its specs. But those that are in the know say things will be interesting this year. So take that as you will...

Will it be an interesting year because of more SM3.0 refreshes or because we are expecting to see D3D10 GPUs in H2?

[BrynS]

With impeccable timing, Digitimes have some brief info from nVIDIA Taiwan's GM supporting an early March launch.

I would agree with Digi that the G71 launch is going to defintately require wide and sustained availability after the constrained GTX 512MB supply and what now seems to be very good X1900 availability.

[Razor1]

Originally Posted by BrynS

I would agree with Digi that the G71 launch is going to defintately require wide and sustained availability after the constrained GTX 512MB supply and what now seems to be very good X1900 availability.

Definilty, othewise they will be crucified.

[digitalwanderer]

Wait a tic, you mean I was right about something?!?!?

[ANova]

I think only one of the rumors is correct, whether that means a 32 pipe part at 500 MHz or 24 pipe part at 700-750 MHz remains the question.

[Carl B]

The rumors for both have come up several times before though. If I had to pick one, I'd go with the 32 pipes vs the ~700 clock; but with 90nm yet to come for their high-end, I see no reason why with this refresh there would or could not be a significant increase in transistor count - the majority of which I would obviously feel be devoted to additional shader quads. I mean right now R580 has almost a full 33% transistor advantage over G70. Something's gotta give besides a clock boost, that's almost certain IMO.

[CJ]

I've posted it once before here that G71 supposedly has a "24-48-24-48 CATZ like the G70". I still haven't completely figured out what they stand for, but here's some more info from the same source:

G70 -> 24-48-24-48 CATZ
G71 -> 24-48-24-48 CATZ
G73 -> 8-24-12-16 CATZ
G74 -> 2-8-4-4 CATZ
G72 -> 2-8-4-4 CATZ

My guess after seeing this was: Colour ops, ALU, Texture Unit, Z/Stencil ops, but that only applies perfectly to G72 and not the G70 since it can only do 16 colour or 32 Z/Stencil ops. And if the second number is the total amount of ALUs, then G73 (GF7600) would be something like an RV530 with 3 ALUs per fragment pipe... So my second thought for the first number was Fragment Pipes, but G72 has 4 Fragment Pipes and 2 ROPs...

So anyone wanna guess again? Cause if G70 and G71 are the same, I'm beginning to doubt if G71 really has 32 fragment shader pipelines and if it's not just a G70 running at 750Mhz thanx to 90nm + lowk. But this goes against what was written in one of the Hexus.beans which said that a G71 needed to be clocked at appr. 430Mhz to reach the same 3DM05 scores as a GTX512. But if we knew what the number 48 has to do with G70 or what the CATZ stand for... this would really clear up a lot about the upcoming G71 architecture.

[digitalwanderer]

Originally Posted by BrynS

With impeccable timing, Digitimes have some brief info from nVIDIA Taiwan's GM supporting an early March launch.

By February 1, a complete lineup of its GeForce 7-series GPU line will all be available worldwide, said Sun.

Nvidia is scheduled to introduce a new GPU series to target the high-end segment in early March

So a month after they have availability on the G7 line they're going to release a refresh?!?

Sorry, the paragraph just confused me a bit.

[BrynS]

Originally Posted by digitalwanderer

So a month after they have availability on the G7 line they're going to release a refresh?!?

Sorry, the paragraph just confused me a bit.

Yeah, apart from being a very short news piece, even for Digitimes, the two statements appear contradictory, unless the February 1 is a typo.

If it isn't an error, then I was thinking it could be a reference to the AGP 7800GS, which has been mentioned a couple of times recently with a launch date of February 2nd. By February, the 7300GS should be available in the US at least and together with the 7800GT/GS/GTX/GTX-512 could be marketed as a complete lineup, although it's still missing the 7600 equivalent, high-end refresh (G71) and any other cut-down, segment fillers.

[Fodder]

I think the February 1 line may be a reference to the 7300, worded terribly.

Edit: Yeah, beaten to the punch.

[Joe DeFuria]

Originally Posted by no-X

...And increasing texturing power without adequate bandwidth could lead to lower-than-expected performance. I can be wrong, of course, and nVidia could be waiting for speedier GDDR4. In this case, the performace could be brilliant, but it's availability and price not.

First we'd have to figure out what "32 pipes" means.

G70 has 16 pixel pipes (ROPS), 24 TMUs, and 48 "pixel shader ALUs". To be an increase over G70, 32 must either relate to ROPS or TMUS. In either case, I would think such a part would be bandwidth limited way before ROP or TMU limited. With "reasonable" RAM, I could see maybe 16/32/48+ (ROP/TMU/ALU) set-up, which may provide some boost to current games.

Unless, of course, G71 includes a 512 bit external bus to memory...then all bets are off.

[Joe DeFuria]

Originally Posted by CJ

G70 -> 24-48-24-48 CATZ
G71 -> 24-48-24-48 CATZ

Well, if G70 and G71 are architecturally different, it doesn't make sense for them to be described the same way. Perhaps there is a typo with G70?

G72 -> 2-8-4-4 CATZ

My guess after seeing this was: Colour ops, ALU, Texture Unit, Z/Stencil ops, but that only applies perfectly to G72 and not the G70...

Actually, to make sense for G72, It would be:
ROPS/Shader ALUS/TMU/Z-Only ops. (Edit: oops...that is exactly what you said!)

To describe G70, it would be:

G70 -> 16/48/24/32

It would also make the G71 a curious part....it would increase the ROPS and Z and leave the shading power untouched...almost directly contrary to the ATI approach.

[Fodder]

Those 48 ALUs can also be referred to as 24 'shader processors' (or whatever your IHV of choice cares to call them), hence the 32 is applicable there as well.

[Joe DeFuria]

Originally Posted by Fodder

Those 48 ALUs can also be referred to as 24 'shader processors' (or whatever your IHV of choice cares to call them), hence the 32 is applicable there as well.

Good point.

[R300King!]

Originally Posted by BrynS

With impeccable timing, Digitimes have some brief info from nVIDIA Taiwan's GM supporting an early March launch.

I would agree with Digi that the G71 launch is going to defintately require wide and sustained availability after the constrained GTX 512MB supply and what now seems to be very good X1900 availability.

He says, "Nvidia is scheduled to introduce a new GPU series to target the high-end segment in early March, according to Paul Sun, general manager of Nvidia Taiwan Sales. By February 1, a complete lineup of its GeForce 7-series GPU line will all be available worldwide, said Sun."

Well, according to this article it seems that he is talking about a new series to be announced, the 8-series maybe, to be announced beginning of March. And by Feb. 1st they will have their 7-series cards(7800GTX 512, 256, mid, low end, etc) all widely available. That's the way I read that.