Geforce NV50 Canceled [Inq]

I heard the NV60 and NV70 have been cancelled too.

 

erased

 

I though that what is the NV40, used to be the NV50. And, looking at NVidia's past history, they've never really made wholesale architectural changes from one generation to the next. It's been more bolt on this, re-arrange that, or update a percentage of its functional units. I'm sure that eases compatibility issues, qualification, driver development, etc. And the current NV40 was likely their biggest departure yet. It's my belief that they had hoped and designed what is the current NV40 with the intention of being their long term platform upon which they would continue to extend.

If MS is moving towards a unified architecture, output from vertex and pixels shaders can easily become input for the other. Building data paths and branch logic for a diverse architecture to support that functionality becomes problematic. NVidia knows what happens when you don't model your hardware closely enough to the API. And while they may still be pressuring MS to not forced unification, and help increase the longevity of their current platform. I think they see the writing on the wall and are going to have to make the move to a unified VS/PS architecture. IMO, of course.

Based on its code name NV47 must be an NV40 derrivative, and as such, is likely to lack any significant architectural changes. It will be interesting to see what they come up with in their effort to maintain their fiscal responsibilities.

 

I just don't see why an HLSL programmer need care about unification. Unification of VS and PS units seems to be mostly a load balancing efficiency/performance issue, and not a fundamental change in the programming model. Some restrictions in SM3.0 may be relaxed in DX-Next, but I don't see this as dictating the underlying implementation. An architecture with separate VS and PS units could still loop back data between stages, without neccessarily sharing the units.

 

My guess is that Inq is misinterperting there sources for info - or possible completely rumors.. Who really knows?

I would think with success of 6 series and fact that next version of direct X is long way off, NVidia could put more time into there designed - or they may have found a totally new way of doing something.

I am curious with release of 6 series, what does this change in ATI vs NVidia users percentages in cards. I would expeect that NVidia has grown.

But that is just my opinion.

 

IT makes sense actually .

The nv50 would have to come out in 2005 .

IT would be in a very bad place at that point .

You'd have the option of dual 6800ultras on one end and then on the other hand lots of longhorn news will be coming with a release in 2006 . Which means many people will want a dx next part .

it may be better off putting the work into making the nv4x tech clock higher .

Not to mention i don't see much faster ram coming out any time soon .

 

Rockster: The NV40 is clearly NV30 based, want it or not. IMO, the NV10->NV20 jump probably was a bigger one. The NV50, while still sharing much from older generations, was AFAIK supposed to be a bigger architectural revamp. But all of this info could be outdated by now I admit.

Uttar

 

Maybe Nvidia knows how R520 performs or will perform

 

Not to mention needing 30 molex power connectors all around the edge of the board.

 

Well, it is conceivable that there are situations where you wouldn't want to do this, however. That is to say, if you have some software that is designed for a unified architecture, it is conceivable that there will be some shaders that just don't entirely fit the pixel shader or vertex shader paradigm very well (i.e. a mixture of the two).

 

In the transition from NV10 to NV20, Nvidia kept the register combiner architecture for the pixel pipeline, added a the vertex shader, and made the tex pipeline more configurable and precise, with fp precision. In the transition from NV30 to NV40, Nvidia did away with combiners and made the pixel pipeline almost equal in flexibility and functionality (although still operating in quads) to the vertex pipeline. I read in a 3DGPU article or interview that NV40 was a clean slate design. Architecturally, the scheduling and parallelism is more advanced, the alu's themselves are superscalar, and the number of pipelines were quadrupled. I'd say that NV40 was as radical a jump from NV30 as NV20 was from NV10.

 

Anyway, I don't think this is such a big loss for nvidia, at least not in financial terms. AFAIK there were not planned any NV5x derived from it; the actual/future Nv4x products were suposed to be used in mid to low end cards.

If Nv50 was such a big project for in terms of actual products, i would have expected NV51 to NV59 variants, and having it placed at a definitive point in API evolution.

To have just one chip placed almost randomly somewhere in between SM3.0 and SM4.0 doesnt look like a product they expected to sell, looks more like a technology show off. and with Longhorn coming sooner than expected they needed a SM4.0 family to sell.

I'd say they took whatever tech they developed for NV50 and decided to make it bullet-proof SM4.0 no waivers

and R400 and NV50 look very similar:

- long time projects that were suposed to be breakthrough
- both almost SM4.0 but not quite.
- both cancelled
- both became the basis for major architectural shifts in technology

I still think Nvidia would have tried to sell it if it would have beat the pants off R520, surprisingly this news comes after R520 tape out and (just guessing) in time for the return of first samples to ATI.

Just how fast that thing is, anyway?

V

 

Maybe it's time somone posted a graph of (estimated) performance based on number of PS quads and vertex shaders. What I'm getting at is the ability of NV40 to scale upwards without any architectural changes. Base the graph on the scaling from 6800nu to 6800 Ultra.

Presuming that NVidia can go with 110nm on a 24 or 32 pipe version of NV40 with, say, 8 vertex shaders, it would be interesting to see what the estimated limits are, using theoretical clock speeds (i.e. around 450-500Mhz core and 700MHz DDR3). Then there's the question of 256-bit bus versus a 512-bit bus (guessing that fall 2005 products would have a 512 bit bus). And why not 90nm by the end of the 2005?

Just curious really whether NVidia thinks that with functionality wrapped up for the next 12-18 months (SM3.0 is effectively complete from NVidia's point of view) ahead of Longhorn that it needs to do anything other than increase parallelism.

Jawed

 

Maybe it's time people started considering this rumor for what it actually is: a baseless rumor. The Inquirer posts baseless crap. No reason to bother listening to them. What's more, at most this just means that nVidia decided to change direction on its next product. Given that we don't actually know what its direction for its next product was, what Fuad said really doesn't change anything important (if it could be considered to be reliable anyway).

 

I think Nvidia has been hanging around with Intel to much....

 

Hey, that's next week's announcement! Turns out a secret codicil of the Intel/NV agreement has NV abandoning the high-end entirely to ATI (which will inevitably result in slowing down high-end performance increases and more expensive high-end parts) in favor of concentrating on making better integrated graphics with Intel chips. SSE5 will only work with Intel/NV graphics chips. . .

But don't tell anyone yet.

 

in fact , there was not any project named as NV50, and NVIDIA will not use NVXX as a project name anymore .

 

R400 wasn't really cancelled according to my understanding. It was more like a major shift in ATI's roadmap, where architectures have been shifted into the future in relative terms (changes applied to the future ventures as much as time and resources allow).

As for either the old R400 and the hypothetical NV50 being even half way SM4.0 I very much doubt it, even more for the first case since I'd speculate that it's quite older. If we'd be talking about NV50 I'd speculate or imagine that conception of the design might have started somewhere in 2003 and by that time it was already pretty clear that WGF (DX-Next back then) won't be around the corner.

Last but not least if there wouldn't be any redundancy involved with a Geometry Shader or the mythical PPP it would had been earlier on roadmaps than it is today. SM4.0 isn't all about unified PS/VS calls. Flip back to the DX-Next article here on B3D and the date it was written. Since the topology part of it has become a name as in Geometry Shader; unify shader calls within a future API and there's more to unify than just PS/VS. You may also add GS to the mix to complete the picture.

Anyway I still can't get rid of the feeling that 2005 will be a very and I mean VERY boring year, but that's just probably me.

 

As simplistic as it may seem, that's an interesting thought and would explain quite a few things. ATI is confusing me with it's codenames and roadmap changes for quite some time now and it's already hard enough to recognize what is what.

 

Is it possible that Nvidia found out some results from the R520 and realised that the NV40 would not be up to scratch (ala NV30 and R300) and so decided to cancel the project.

If so.. how far is the NV55 from being released?

Again .. if the naming convention for the NV50 has changed .. it could be the reason for the "supposed dropping of the NV50".

Will the NV47 be up to scratch to challenge the R520?

US