NVIDIA: Beyond G80...

[kyetech]

OK, I've been trying to think how and when nvidia will refresh the top end product. It seems that around feb time there will be an 80nm g80 with minor tweaks and a small speed boost but then what?

Heres my complete guess, could we expect something like this?

Feb 2007 80nm : Small speed boost // gddr4 ?
Fall 2007 65nm : 1GB mem // 512 bit memory // 32 rops // 800 million transistors // speed bump
Spring 2008 65nm : GX2 stylee g8x derivative ?? or is a single bigger die more likely this time? And if it is a single bigger die, then what do you think they would attempt for the specs, given the process constraints of 65nm?
summer / fall 2008, 55nm half node G90, with a rumoured 'theinq' style r700 distributed multi chip design?

I'd be really interested in seeing how people think it will pan out, since chip designs and board layouts are getting more unorthodox and more difficult to guess than patterns of refreshes in years gone by?

I know megadrive1988 tends to like these speculative what does the future hold type threads. I'd be interested in hearing all your ideas.

 

[Arun]

The following is speculative, and should not be considered based on any non-public information, as it quite simply is not so!
The codenames used are most certainly incorrect, and some might not even have their own codenames. As such, they are used exclusively to permit further discussion, and nothing else!

February->March 2007
G81: Optical shrink of G80 on 80GT (G80 currently is on 90GT - R600 is either on 80GT or 80HS). Introduces the 7/8 clusters SKU. Exact specifications (identical to G80 on 90GT or not; GDDR4 or not; etc.) depend on the R600's specifications. 1.1GHz GDDR4 and 600-625MHz is possible, if considered necessary. If such a step is taken, the parts with redundancy would likely also use GDDR4, which would define the 8900 line-up. There will NOT be a notebook version.
G82: G8x on 80GT with 6 native clusters and 5 native ROP partitions. Can act as a 8800GTS (which will then cost $399) or a 8700-series GPU, which would likely have 5 active clusters, 4 active ROP partitions and a 256-bit memory bus. This is needed for the 8800(/8900?)GTS, as G81 will mostly be 8/7 cluster parts due to improved yields with the smaller 80nm die. There will be a notebook version. Maybe pin-compatible with G81(?)
G83: G8x on 80GT(?) with 4 native clusters and 2 native ROP partitions. 192-bit memory bus and roughly $199 target introduction price. 8600/8500-Series, possible version with 3 active clusters.

March->May 2007
G84: G8x on 80GT(?) with 3 native clusters and 2 native ROP partitions. Version with 2 active clusters and 2 native ROP partitions. 128-bit memory bus. 8400/8300-Series. Maybe pin-compatible with G83(?)
G8i: G8x-based Intel IGP on 80GT(?) with 1 native cluster and 1 ROP partition, that has even less blending units, maybe no extra double-z, etc.(?)
G8a: Same as G8i, but for AMD's socket AM2.

June->August 2007
G85: G8x on 65nm, 2 native clusters and 1 native ROP partition. Sold at ~$99. Finally replaces G73's 80nm shrink. G72's 65nm shrink will already have been released for some time, and will remain in heavy production for the ultra-low-end SKUs.

September->December 2007
G90:
65nm, Q4 2007; 400mm2+
1.4-1.6GHz GDDR4, 384-bit Bus
1.45-1.75GHz Shader Core Clock
6 ROP partitions, but beefier
625-675MHz Core Clock
FP64 Support (Slow!)
32 MADDs/Cluster
24 Interps/Cluster
10 Clusters
Q1->Q2 2008
G91: 7(?) native clusters, 65nm
G92: 4(?) native clusters, 65nm
Q2->Q3 2008
G9I/A: 2(?) native clusters IGPs, 55nm(?) (competes with Fusion?)
G93: 3(?) native clusters, 55nm(?); replaces G85
Q4 2008
G94: 7(?) native clusters, 55nm(?)
G95: 12(?) native clusters, 55nm(?); 9900-Series
G96: 4(?) native clusters, 45nm(?); G93 moves into the ultra-low-end
Q2 2009
G100: ...

BTW, in terms of Quad-SLI, I think you're much more likely to see those with G82 and G91 than with G81 or G90. That's still a hefty performance gain, and would be much more reasonable in terms of power consumption.


Uttar

 

[Skrying]

I strongly believe that no version of G80 will have a 512-bit memory bus. IMO the idea of one having a 512-bit bus is nuts. Is it truly possible that they'd move to a 512-bit with a smaller die? I think not.

Uttar's prediction seems about where I'd put everything as well.

 

[Lezmaka]

How likely is it that the clusters could have different number of ALUs (more or less, depending on positioning)? Or that the bus width per ROP partition is halved for the lower end?

 

[Razor1]

I don't think nV will go with more ALU's per cluster as that would increase thier batch size, more likely they would increase number of clusters, and the way thier bus is split amoung the 6 clusters right now is 64 bits each, it would be advantagous to go this route.

 

[dnavas]

I thought the same as Razor when I saw Uttar's original prediction of 32ALUs/cluster. I still feel it's more likely that we'll see more clusters than ALUs. Another possibility is to leave the cluster count as is for purposes of interaction with, say, TEX units and external L2, but have two schedulers and parallel SIMD sets of ALUs. You might also double texture addressing units, which would eliminate "free" filtering, but allow better TEX sharing between parallel sets.

 

[Rufus]

I don't think nV will go with more ALU's per cluster as that would increase thier batch size, more likely they would increase number of clusters, and the way thier bus is split amoung the 6 clusters right now is 64 bits each, it would be advantagous to go this route.

Also nV must have done some sort of semi-custom design in order to get the ALU cluster to hit 1.35Ghz. Being able to drop-in identical clusters across different chips should save a huge amount of engineering work. I highly doubt anything about the clusters (except bug-fixes, which we probably won't hear about anyway) will change untill at least G90.

 

[MulciberXP]

I'll throw my prediction in that we wont see G90 until 2008.

 

[INKster]

I'll throw my prediction in that we wont see G90 until 2008.

Based on previous NV's experiences, that's not entirely liquid.

- NV40 (April 2004) -> G70 (June 2005) -> G71 (March 2006)

A zero at the end of the codename does not grant an automatic architecture step-up, like if it was a brand new generation.
But i agree that the "1 year from now" timetable is a little too tight for a G90, even with Jen-Hsun's comments about a load of codenames to come.

 

[Sunday]

NVIDIA will advance with the pace of TSMC’s advancement.

NVIDIA doesn’t have need to expand the G80 with more SP’s. Arch is excelent, and almost certainly AGP (AMD Graphic Product Division) will not succeed in de-crowning the NVIDIA in the performance arena with R600).

What they need is the GPU that is smaller so that they can achieve same record shattering revenue records in the future.

So they will use 80nm (spring 2K7), 65nm (fall 2K7), and 55nm (summer or fall 2K8) manufacturing technologies, to make smaller, and more profitable GPU derivates from the G80. They will scale performance with more clock, and GDDR4!

In 2008, they will introduce G90 on 55nm with architecture advancements toward DX10.1 compatibility. There’s high possibility that G90 will use 512bit memory bus interface!

 

[Skrying]

NVIDIA doesn’t have need to expand the G80 with more SP’s. Arch is excelent, and almost certainly AGP (AMD Graphic Product Division) will not succeed in de-crowning the NVIDIA in the performance arena with R600).

How are you so certain about that? From what the rumors its been that the R600 will be faster than G80.

 

[INKster]

In 2008, they will introduce G90 on 55nm with architecture advancements toward DX10.1 compatibility. There’s high possibility that G90 will use 512bit memory bus interface!

But DX10.1 is supposed to be out in one year along with Vista's service pack 1 (late 2007), according to every reference i have read in the last few months.
Would Nvidia miss the chance to add those changes earlier on a fall refresh part, just to give ATI the possibility to release a R600 refresh with DX10.1 capabilities during the crucial holiday season ?

That is, of course, unless you know something about DX10.1's timetable that we don't...

 

[Sunday]

In mine opinion ATI has made wrong bet with complicating the shader processors with vector processing capability. NVIDIA was able to achieve such a high frequencies thanx to the simplicity of their scalar SP’s.

There is no way that ATI will go beyond present 650-700 MHz with similar design of the shader pipes with R580! Fact that they A1 silicon didn’t even worked, and that is even bigger then G80 is drawing pessimistic picture in mine mind.

So yeah, I’m quite convinced that R600 will not surpass G80!

And let’s not forget that R600 is 80nm GPU (I repeat it’s bigger then G80), and that NVIDIA has that card in their sleeves to use!

 

[Skrying]

Ummm, I'm fairly certain you're understanding of both what R600 will be and what G80 is, is highly mixed up.

 

[BRiT]

Why must every Nvidia thread derail into an argument around AMD/ATI R600 ?

 

[INKster]

Why must every Nvidia thread derail into an argument around AMD/ATI R600 ?

Causality
One company's actions depend on what the other is doing.

 

[Razor1]

There is no way that ATI will go beyond present 650-700 MHz with similar design of the shader pipes with R580! Fact that they A1 silicon didn’t even worked, and that is even bigger then G80 is drawing pessimistic picture in mine mind.

And let’s not forget that R600 is 80nm GPU (I repeat it’s bigger then G80), and that NVIDIA has that card in their sleeves to use!

Yeah thats what I heard a couple of weeks ago, but nothing really tangible, or confirmed yet, but this makes two of us hearing the same thing. But I think ATi is shooting for 750 mhz. Now I didn't get any info on the number of ALU's or configuration.........,

Its running hot 20-50 watts above the g80, its a huge chip 750+ million tranis.

 

[silent_guy]

In mine opinion ATI has made wrong bet with complicating the shader processors with vector processing capability. NVIDIA was able to achieve such a high frequencies thanx to the simplicity of their scalar SP’s.

Despite much higher complexity of their execution units and multiple parallel ALU's, CPU's have been running at speeds above 3 GHz for years.
Why? Because Intel and AMD decided that was necessary to do. So it became as a necessary design goal. So the engineers spent their time (a lot of it!) to make it that way.

It's no different for a GPU. If a GPU shader is not running a 3 GHz, it's because somebody decided, after careful analysis of all pro's and con's, that it was not worth doing. That's all there is to it. It has very little to do with the execution unit being scalar or vector oriented.

 

[Geeforcer]

Well, CPUs have been stuck at 3GHz for years now, and no one seems to mind.

 

[KimB]

Well, I think the first change is that they'll beef up the first MUL unit, that currently appears to only work with interpolation, so that it can do a full MAD. The G8x, as it stands today, is so damned good that I expect it to last 3-4 years with only minor tweaks and die shrinks. I do expect the refresh part (which will probably be labeled "G90") to be noticeably more efficient, from the lessons learned with the G80, as well as having more functional units.

What nVidia does beyond that, well, I expect will depend in part upon what ATI does with the R600.