Geforce 8900: an overclocked G80 that should give R600 a run for the money.

[sladesurfer]

The title says it all, really. NVIDIA is planning on overclocking the G80 as far as they can get it, which alone wouldn't give you too much a performance increase. However, NVIDIA also has a figurative ace-in-the-hole, a driver tweak. What exactly this driver tweak does is unknown. Whatever it is, though, will give ATI some serious competition when they release their gigantic R600.

http://www.techpowerup.com/?25258

[GrimReaper]

lol 8900 overclocked G80 testing manufacturers RMA ?
7900 anyone

[xlink]

by driver tweak I assume they mean GDDR4 memory and a moderate OC...

because people like use would just crack the drivers...

[nn_step]

nice linkage, linking to a page that uses the Inq as a source. Can't say I believe it completely.

[HamidFULL]

I remember unborn 7800 Ultra that known as 7800GTX 512MB ...
here would be 8800 Ultra ...

[milkcafe]

the upgrading game between nVidia/ATI wiil return like this

[G.Foyle]

One of the comments to original article:

i can guess what the driver tweak will be maybe it will be a working driver :roll:

[turtle]

I still think they're holding off transition to a new process until 65nm.

Who wants to bet a cookie that the driver tweak activates the 'missing mul' for general shading?

G80's architecture is dual-issue scaler + MUL, yet the MUL is never issued. I don't know if it'll add (in some cases) 1/3 performance to G80, but it'll do something in the positive. I always figured this driver tweak would be for the 8800 when R600 is released. If they do not activate it on G80, but release the same silicon with it turned on and call it an 8900, I'd be royally pissed if I owned an 8800.

About the driver tweak being one that actually makes the G80/SLI fully work with Vista, I find that rather amusing.

[BlackX]

One of the comments to original article:

Its written "overclocked G80" so I wouldnt expect GDDR4 or 80nm process. Im a bit disappointed because I was really hoping to see G81 on R600 arrival. I cant see why anybody would buy overclocked and probably overpriced G80.

[zakelwe]

The MUL theory is interesting. Other possibles apart from this and an overclock are whether the 8800GTX has all it's stream processors enabled to aid yields or whether another 32 ( pushing it up to 160 ) can be used in a pinch for a low volume very expensive part. That would assume some built in redundancy which of course is just a wild guess....

You have to assume there is something here or there abouts to counter R600 which has to be quicker to stop the "ho-hum" factor.

Of course the real counter would be to drop the 8800GTX price massively but they ain't going to do that unfortunately I think

Regards

Andy

[Tonucci]

I was hoping to see 1Gb GDDR4, 512bit in the refresh :P Im asking too much i guess...

[onewingedangel]

If its an overclocked g80 it should be called 8800ultra or 8850. After all its just a high binned variant much like the X850XTPE was - a low volume part just so that they can appear competitive with the competition - they don't even have to ship that many as long as the benchmarks are out there. Most people would still go for the gt or gtx anyway if its only a matter of clockspeed.

The g81 should get the 8900 tag.

[Starscream]

wel a while ago there was the rumor that Nvidia ordered 80nm G80 cores at the fab.

COuld be those 80nm cores r meant for the refresh.

like others i doubt this refresh will get the name 8900 if it is JUST a a new process.

i guess its just something to compete the R600 and that they continue work on the G81 to fight the potential X2900.

[vowkilmer]

Its written "overclocked G80" so I wouldnt expect GDDR4 or 80nm process. Im a bit disappointed because I was really hoping to see G81 on R600 arrival. I cant see why anybody would buy overclocked and probably overpriced G80.

My thoughts exactly.

[mr_knowitall15]

who didnt see this coming? Theyve had like a 6 month lead for some R&D to improve the g80 so thay can counter or beeotchslap R600 whenever it gets here.

[Anemone]

I would expect at least 80nm but better to just go for 65 if the fabs can make good yields and they get a working design. (those are big IF's I do realize). GDDR4 would help, fight back the superior 512 memory controller on the 600. But there is little chance they'd get a redesigned controller.

As for added activation of features, that's normal for Nvidia. In fact there have been several first gen vs refresh feature issues in many prior years. So I'd expect them to do it again... Yes that doesn't bode well for the first adopter squad. But it is very normal for Nvidia to do.

That 6 month lead time hasn't been enough for them to get a quality Vista driver, so I wouldn't hold my breath. They may very well just give the lead over to the 600 and let that be that, and worry about a revised chip in a few months. If you look at the stock vs the overclocked scores I think you'll find the clues. A mildly overclocked G80 "ultra" won't do it against the 600. They will need a solid 20-25% boost to hold the lead solidly. Even then they may not hold it 100% in every bench.

Have to see. But 65 nm and faster GDDR4 should be a minimum to see the 8900 moniker. Things short of that will just be premium binned "ultra" parts and ho hum on those.

[arisythila]

LOL... Werd™

[Cybercat]

I still think they're holding off transition to a new process until 65nm.

Who wants to bet a cookie that the driver tweak activates the 'missing mul' for general shading?

G80's architecture is dual-issue scaler + MUL, yet the MUL is never issued. I don't know if it'll add (in some cases) 1/3 performance to G80, but it'll do something in the positive. I always figured this driver tweak would be for the 8800 when R600 is released. If they do not activate it on G80, but release the same silicon with it turned on and call it an 8900, I'd be royally pissed if I owned an 8800.

I was thinking about that actually. As you said, it's unknown if it will really give significant improvements, a MUL has less functionality than a MADD, it's not such a linear improvement as theoretical numbers make it out to be.

I've read many highly technical threads on the G80, and to this day I still don't know how it's able to perform so well without that MUL. We're talking a highly efficient chip here (mmmm.....scalar). The only way I can figure is if you go by what some people in the industry have said, the G80 is closer to 90% utilization compared to about 60% of last generation GF7 architectures (I wish I had a link), plus about 40% more raw theoretical shading power than the best GF7 card (not including GX2), put that together that's 70% improvement at best. But that results from a lot of speculation. The 8800GTX still ends up nearly doubling 7900GTX performance, for which I guess added vertex and texture performance can be attributed.

[turtle]

I was thinking about that actually. As you said, it's unknown if it will really give significant improvements, a MUL has less functionality than a MADD, it's not such a linear improvement as theoretical numbers make it out to be.

I've read many highly technical threads on the G80, and to this day I still don't know how it's able to perform so well without that MUL. We're talking a highly efficient chip here (mmmm.....scalar). The only way I can figure is if you go by what some people in the industry have said, the G80 is closer to 90% utilization compared to about 60% of last generation GF7 architectures (I wish I had a link), plus about 40% more raw theoretical shading power than the best GF7 card (not including GX2), put that together that's 70% improvement at best. But that results from a lot of speculation. The 8800GTX still ends up nearly doubling 7900GTX performance, for which I guess added vertex and texture performance can be attributed.

All good points on possible reasoning. Although TMU's may definitely have something to do with it, as surely that butchered R580, I don't think the vertex performance does to that large of an extent.

I too question how it performs so well when the GFLOP number, for instance, is substantially lower without the MUL. It actually puts it more on the level of a x1900, which makes it all the more amazing. True enough, G80 and a scaler architecture in general is much much much (much) more efficient than R580's (and of course G70's) setup, but still, can that alone justify the difference? I suppose the MUL is working, as B3D points out, doing 'interpolation and perspective correction', which I imagine also has something to do with it's efficiency (as I think people overlook the MUL is actually doing something), but I imagine the driver tweak to allow it to do general shading will allow for an improvement in overall performance in at least some scenarios,if not all. How much...Well, that's the question. Perhaps it will bring the over-all efficiency down when looking at it from a grand scheme because, as you said, a MUL is not as useful as say adding another scaler unit (MADD), but if it's there and can be used more efficiently in general shading than what it's doing now in some (or all) scenarios than doing what it is now, I imagine they'll use it for that. I truly do believe that is what the 'driver tweak' is; this is the ace up their sleeve. The question to me though, again, is 'How much will it matter?' Apparently Teh Inq, or perhaps nvidia, thinks quite a bit.

[Daveb2012]

http://www.techpowerup.com/?25258

this doesnt sound promosing, its going to have a "special" driver Its probably just going to increase the clocks, voltage, and likeliness to fail.

[Nanometer]

Special driver? Sounds stupid and a waste of time. I think this GPU will hold me for a long time.

[BlackX]

Special driver? Sounds stupid and a waste of time. I think this GPU will hold me for a long time.

Sounds like stupid commercial tricks IMO.

[Cybercat]

All good points on possible reasoning. Although TMU's may definitely have something to do with it, as surely that butchered R580, I don't think the vertex performance does to that large of an extent.

I too question how it performs so well when the GFLOP number, for instance, is substantially lower without the MUL. It actually puts it more on the level of a x1900, which makes it all the more amazing. True enough, G80 and a scaler architecture in general is much much much (much) more efficient than R580's (and of course G70's) setup, but still, can that alone justify the difference? I suppose the MUL is working, as B3D points out, doing 'interpolation and perspective correction', which I imagine also has something to do with it's efficiency (as I think people overlook the MUL is actually doing something), but I imagine the driver tweak to allow it to do general shading will allow for an improvement in overall performance in at least some scenarios,if not all. How much...Well, that's the question. Perhaps it will bring the over-all efficiency down when looking at it from a grand scheme because, as you said, a MUL is not as useful as say adding another scaler unit (MADD), but if it's there and can be used more efficiently in general shading than what it's doing now in some (or all) scenarios than doing what it is now, I imagine they'll use it for that. I truly do believe that is what the 'driver tweak' is; this is the ace up their sleeve. The question to me though, again, is 'How much will it matter?' Apparently Teh Inq, or perhaps nvidia, thinks quite a bit.

A unit dedicated to perspective correction isn't even worth mentioning in the white papers, much less including in theoretical FLOP numbers, as many architectures particularly the R300 architecture have a dedicated perspective correction unit running in the background that ATI never bothered to include in the shader diagrams or in theoretical numbers when describing it to the press. It's essentially free in old architectures, suddenly by illuminating the MUL in the G80 and saying it's needed to come up with the theoretical shader numbers the G80 is capable of, NVIDIA is saying next gen means age-old routines like perspective correction are no longer "free". But in marketing, 518 GFLOPs looks a hell of a lot better than 345 GFLOPs.

But this move by NVIDIA brings up an interesting take on what their opinion may be of the R600. It may show a lack of NVIDIA's confidence in the efficiency of the R600 as compared to the G80. There's a very good chance the R600 will use the same 4-way Vector shader structure of the R300 ancestry, which is a pretty tried and true architecture, but will prove dismal compared to a fully scalar architecture that NVIDIA is boasting. They'll have to use a mass amount of these shaders and a high clockspeed to create shader performance on par with NVIDIA. On paper, the theoretical numbers will actually be slightly higher (suggesting a clockspeed of slightly over 800MHz holds true), but in real world utilization NVIDIA will have the upper hand, which is crucial. Instead, the R600 will have to rely on other aspects to obtain the win over the G80, and that is raw fillrates, both pixel and texel, and memory bandwidth. That will give ATI a much needed boost in heavily bump-mapped scenarios, HDR and post-processing effects, anisotropic filtering, and anti-aliasing. Add all those together and they make up the bulk of what a graphics card has to do in modern games, and thus the R600 will have the edge.


At least over the first generation of G80.

[ahmad]

Nvidia had been going for a raw performer, not a feature oriented design. G80 mimicks that. If you ask me, if it wasn't for AVIVO, nvidia would have never cared to improve their decoding hardware/software.

I sure hope Nvidia doesn't go out of their way to try and optimize their drivers some more by reducing the visual quality of the g80's output just for better FPS... If they do that who knows how long it will take me to recover from that stupidity.

[turtle]

A unit dedicated to perspective correction isn't even worth mentioning in the white papers, much less including in theoretical FLOP numbers, as many architectures particularly the R300 architecture have a dedicated perspective correction unit running in the background that ATI never bothered to include in the shader diagrams or in theoretical numbers when describing it to the press. It's essentially free in old architectures, suddenly by illuminating the MUL in the G80 and saying it's needed to come up with the theoretical shader numbers the G80 is capable of, NVIDIA is saying next gen means age-old routines like perspective correction are no longer "free". But in marketing, 518 GFLOPs looks a hell of a lot better than 345 GFLOPs.

But this move by NVIDIA brings up an interesting take on what their opinion may be of the R600. It may show a lack of NVIDIA's confidence in the efficiency of the R600 as compared to the G80. There's a very good chance the R600 will use the same 4-way Vector shader structure of the R300 ancestry, which is a pretty tried and true architecture, but will prove dismal compared to a fully scalar architecture that NVIDIA is boasting. They'll have to use a mass amount of these shaders and a high clockspeed to create shader performance on par with NVIDIA. On paper, the theoretical numbers will actually be slightly higher (suggesting a clockspeed of slightly over 800MHz holds true), but in real world utilization NVIDIA will have the upper hand, which is crucial. Instead, the R600 will have to rely on other aspects to obtain the win over the G80, and that is raw fillrates, both pixel and texel, and memory bandwidth. That will give ATI a much needed boost in heavily bump-mapped scenarios, HDR and post-processing effects, anisotropic filtering, and anti-aliasing. Add all those together and they make up the bulk of what a graphics card has to do in modern games, and thus the R600 will have the edge.


At least over the first generation of G80.

Very well put, and I agree on almost everything.

If the tasks the MUL is currently working on are essentially free or were in past generations with significantly less power, and added to G80's spec just for a >500GFLOP 'checkbox' (there's a Geo line for ya), would it really require 173 Gflops for those tasks? I wouldn't think so. They've got to be cooking up some way to use it. It seems like such overkill for such simple operations.