The G92 Architecture Rumours & Speculation Thread

[Jawed]

Why should NV give this advantage of their architecture up?

Because it costs a lot of die space?

G8x design links MCs to ROPs and L2 (texture + render target?). So for each extra MC there are also 4 extra ROPs and cache (10s of KB it seems). That's a fair amount of die space. I think it's pretty neat arrangement - these things should scale together.

NVidia went 384-bit as an interim step before heading off down the GX2-style road (2x 256-bit). ATI went 512-bit before heading off down the 2x 256-bit road (even if it seems we're likely to see a single MCM with a "512-bit memory interface" when R700 turns up...).

Jawed

 

[3dilettante]

What AA bug is Charlie talking about?

I believe at least two folks from AMD have stated in this forum that there is no bug with AA.

I know Dave has said as much, and I think Sireric may have said AA is fine as well.

 

[Unknown Soldier]

I still say, if the 8800 GT is as fast as the GTX in real world games and comes in at the handsome price($249+) then i'll sell my GTS 320MB to get the GT 512MB.

Reason? Well I have a monitor that does 1680x1050 and while the GTS runs most games at that res. without problems, it does have a problem with DIRT with 4xAA(the AA I like to play my games at) at that res. So if the 512MB does sort out the memory problem(High res and AA) is the cause the DIRT doesn't work and I think, is the issue with DX10 games not working at that res.(since DX10 requires 4xAA on at all times).

p.s. I'll most probably switch over to Vista too.

Hell .. i'd buy a 1Gig GT if it was as quick as the GTX and came at the Price of $299.

US

 

[AnarchX]

Because it costs a lot of die space?[...]NVidia went 384-bit as an interim step before heading off down the GX2-style road (2x 256-bit).

Worry about costs of die-space of a 384Bit MC and then talking about a GX2-solution, which will have a massive amount of redundancy?

And again:
Why should NV advertise 3-Way-SLI, when upcoming high-end-solution does not will be supported? Or do you really believe in 6-Way-AFR?
Why should NV cripple a 290/330mm² Die with a 256Bit MC?

 

[Geeforcer]

No, Charlie has been an NV hater for over 1 year now and to date has never reported 1 bad thing about ATI.

You mean 8600s did not have "broken 2D clock, launch to be delayed by months" and "R600 and Core2" were not the "ultimate gaming platform" of the 2007?

 

[fellix]

OK, let's count them:

 

[Arty]

184.

 

[AnarchX]

What do you think about 754M transistors?

 

[fellix]

Better there is some eDRAM...

...or at least some sort of huge multi-level z/stencil buffer.

 

[Jawed]

Worry about costs of die-space of a 384Bit MC and then talking about a GX2-solution, which will have a massive amount of redundancy?

Eh? GX2 is a premium product.

And again:
Why should NV advertise 3-Way-SLI, when upcoming high-end-solution does not will be supported? Or do you really believe in 6-Way-AFR?
Why should NV cripple a 290/330mm² Die with a 256Bit MC?

Why would NVidia make an even larger die when an 8 cluster version of G92 could be in the region of 80-90% of 8800GTX across the board (and be faster on certain games).

The corner case against G92 is 2560x1600 gaming with AA - that's what GX2 will be for, by the look of things. For a real upgrade, 8800GTX owners will be looking for something that's considerably faster. A 384-bit G92 wouldn't be a compelling product. Whereas a $500 GX2 will be (unless you think 7900GX2/7950GX2 was a terrible product and you're never going there again).

3-way SLI looks like it might be a Christmas campaign (stop-gap), to get 8800GTX owners to spend more money (or get people who have the deep pockets to spend on more than just 1 or 2 8800GTXs before Christmas). Then 8800GTX stops being made and a new horizon dawns in January with GX2.

Can (will) a GX2 work in SLI with a non-GX2 board (i.e. 3-way)?... Is 4-way with 2 GX2s impossible?

Jawed

 

[AnarchX]

Eh? GX2 is a premium product.

Sure, but Jen-Hsun Huang want his high margins, which would not give 2x289mm²-die+2x1GB card...

Why would NVidia make an even larger die when an 8 cluster version of G92 could be in the region of 80-90% of 8800GTX across the board (and be faster on certain games).

Why should NV make a 289mm²/754M die without a 384Bit MC?
A 6C-part with ~200-250mm² @ higher clocks would make here more sense and also in connection with a GX2.

3-way SLI looks like it might be a Christmas campaign (stop-gap), to get 8800GTX owners to spend more money (or get people who have the deep pockets to spend on more than just 1 or 2 8800GTXs before Christmas).

I do not think. 3-Way is the most cleverest solution untill we have something better than AFR or Now-SFR.
And 790i, 780i's successor with 3x PEG Gen2 is announced for Q1 2008.

Can (will) a GX2 work in SLI with a non-GX2 board (i.e. 3-way)?...

Possible, some Chinese run 7950GX2 + 7900GT 512MB in 3-Way-AFR in 2006...

Is 4-way with 2 GX2s impossible?

Sure it is possible, but GX2 was more a temporary solution against R580, with a high margin ~200mm² G71.

4-Way-AFR never worked under D3D and Q-SLI is not supported under Vista until now.

 

[Sound_Card]

That is one big freaking die.

 

[BrynS]

Any word on mobile plans for G92 in high-end gaming/DTR notebooks?

Looking back there does appear to be a precedent for a notebook GPU of this size (GeForce Go 7800 [GTX] at 334mm2) but I don't recall it being particularly prevalent before it was usurped by the much smaller G71 (GeForce Go 7900) at 196mm2. R520 (288mm2) also featured in a handful of DTRs (Fujitsu-Siemens being the only one to spring to mind), but the various mobile G71 iterations have pretty much dominated the notebook gaming field for nearly the last two years.

Unless NV have some more er.. cards up their sleeve it looks as though the gaming notebook market is swinging AMD's way in 2008 with RV670.

 

[NonNative]

Hmm its seem no one talking about new 8800GTS SKU at all ?
Do anyone know about it ? I only know that its based on G92 as well.

 

[AnarchX]

Hmm its seem no one talking about new 8800GTS SKU at all ?
Do anyone know about it ? I only know that its based on G92 as well.

I heard something about dual-slot, $305, 512MB/256Bit, 128SPs and launch at 19.11., but no idea how reliable.

 

[Twinkie]

Would it be safe to assume that G92 is 384bit bus since the transistor count is rumoured to be at 754million? G80 has 681 million transistors plus NVIO. So 73 million are left for both NVIO and VP2 which sounds reasonable.

With this kind of chip, the whole concept of "GX2" makes no logical sense business and technological wise. Wouldn't they have opted for a native 6C part? Is this one of those smoke screens?

Waiting eagerly for the full fledged G92..

 

[Jawed]

Sure, but Jen-Hsun Huang want his high margins, which would not give 2x289mm²-die+2x1GB card...

I'll let the yield&margin zealots argue that point with you...

Why should NV make a 289mm²/754M die without a 384Bit MC?
A 6C-part with ~200-250mm² @ higher clocks would make here more sense and also in connection with a GX2.

It seems you might just as well ask why G80's true successor, the 512-bit 192-SP, 2.4GHz SP clock monster isn't releasing on Monday.

I do not think. 3-Way is the most cleverest solution untill we have something better than AFR or Now-SFR.
And 790i, 780i's successor with 3x PEG Gen2 is announced for Q1 2008.

As late as that? I thought it was coming in November.

Jawed

 

[AnarchX]

It seems you might just as well ask why G80's true successor, the 512-bit 192-SP, 2.4GHz SP clock monster isn't releasing on Monday.

Was not G71 G70's successor?
Why have to have a 12-months-later-refresh more units, also with the situation that competition is very weak?

I think after that we know that G92 is a huge 290mm² or even more die, with like I heard 754M transistors, anything other than a G80@65nm with the known added stuff would be nonsense, like also the GX2-rumors.

And now I think this slides pictures G92, which is supposed to support SPOC:
http://img261.imageshack.us/img261/8684/nvspoctq4.jpg
-> 8 Cluster, 6 ROP-Partitions and a VP2...

As late as that? I thought it was coming in November.

In November will come 780i(C72 = trough BR04 pimped up 680i) and 780a(based on MCP72). The "real" 3-Way-SLI-chipset for Intel C73 aka 790i will come in Q1 08.

 

[Twinkie]

It seems you might just as well ask why G80's true successor, the 512-bit 192-SP, 2.4GHz SP clock monster isn't releasing on Monday.

Because nVIDIA doesn't want to shoot its own foot?

 

[Jawed]

Because nVIDIA doesn't want to shoot its own foot?

Hence GX2...

Jawed