NVIDIA CineFX Architecture (NV30)

edited: Also there are three other things to consider:
-The hand tweaked design of R300 (how much it really improved it?)
-The extra power beyond AGP 2.0 specs.
-Memory bandwith (128bits? , DDR-II?)

 

OT a little: Can someone explain me what's behind the "hand tweaked design" concept? Thx

 

ATi said they took 'an Intel like approach' when actually doing the layout of the chip by hand tweaking it to attain the speeds that they currently are. Normally this process would be analysed and processed using design tools on computer - this is usually the quickest method of getting a design that will opearate but may not necessarily be optimal. Hand tweaking a chip of 107M transistor could take years to finalise (taking into account errors!) so I doubt they 'tweaked' all that much, just a few critical paths.

 

Usually with designs as complex as 100 millions transistors with small time to market and low cost of design people use a lot of automated tools do to a lot of different things like cell design, cells placement and routing.

One cell can be made by few transistors to as much as many thousands transistors.

A hand tweaked design mean that something was done by the designers that could not be done so well by the automated tools. What was hand tweaked I dont know, the only reference about it is this Anandtech page:
http://www.anandtech.com/video/showdoc.html?i=1656&p=2

 

Thx you Dave and Pascal, that's is why i like this forum. Even if i'm not a 3D specialist, i can learn some interesting things from you.

*Long life to B3D*

/end OT

 

Regarding the R300's 107M transistor and the high clock speed 300+ MHz(over Matrox etc):

DX9 could well introduce deeper pipelines or ATI could just have decided to make them deeper (aka AMD vs Intel's approach with Athlon and Pentium 4).

I guess you could also add more non-functional transitors here and there to make sure that the chip doesn't get too hot on some specific spots on the die (although I don't think they had silicon to burn! ).

 

Technically incorrect. How can you say Quake3 uses little to no hardware geometry compression? Or did you mean to say Quake3 benefits very little from hardware geometry acceleration?

As for JF_Aiden_Pryde's question to me - physics and AI account for the majority of the mentioned-games' CPU dependency. And I would assume you ask simply because you don't own either of the games... shows exactly what I meant in my previous post... there is no truth in the saying "Quake3 is relevant for benchmarking video cards since it scales so well with better video cards even if it's old because there are a few other games using its engine."

You want better games, better looking games or tech demos when reading a review at Beyond3D?

 

The nice thing about Q3, was you could overclock card A by an amount X. And expect the next data points to be linear as well (that is, by an amount cX).

That is, it scaled with memory/core clock in a nice, easily identifiable way which was convenient to benchmark. Even on fast computers, this seemed to be the case until not so long ago.

Where it has lost its edge really, is in the the demands it placed on multitexturing, its relatively low overdraw, and its applicability for heavy transform limited cases.

Its still a good test, but not nearly as useful as it once was. Sites that use it in exclusivity are missing the point on an array of different scenarios.

 

Actually, rev, what he said was "Quake3 uses little to no geometry ACCELERATION"...
nothing about compression.
he's totally correct.

 

BTW is it just me did anyone realize with every new core from ATI we see one less TMU per pipeline from R100-R300 I wonder if they won't have any in their R400 cores.

 

I also agree that it will be hard, unless Nvidia does something we don't expect, like release an SLI enabled (2 chips) NV30 reference board and throw it at the big web sites...

Whamo! Top of the performance charts! The darling of the enthusiasts!

Of course the cost would be crazy, and manufacturers would probably only make single chip versions in the begining, but Nvidia would steal all the headlines ...

Just a wacky idea...

Guenther

 

I would love to see such a beast, even if I couldn't afford one.

Then again, with all the new purchases that we're making for the new house, I could probably bury the cost of a $600 graphics card in among the furniture & lawn equipment.... 8)

 

I've actually gotten my wife hooked on the idea of a theater room with a front projector. All it means is mac'n'cheese for a year or so.

 

Yeah....a home theatre is tentatively in the plans for us...but then we found out that to put a damn fence around the back yard like we want will probably run well over 10K.....

Sigh....and people think graphics cards are expensive.

 

You should try plumbing. We have over $6000 in plumbing FIXTURES, not even counting the pipe in the wall.

Who knew an f'n sink cost $500

 

Let's look at what we know:

1. In features, we know of nothing that the R300 supports that the NV30 does not, and the NV30 certainly supports more.

2. In performance, the .13 micron process should allow for superior performance. The available memory bandwidth is the great unknown here.

3. In FSAA performance, nVidia already does have MSAA, so it shouldn't be too hard for nVidia to suceed here.

4. In FSAA image quality, nVidia would need to change their design, but this shouldn't be too terribly hard.

5. In anisotropic, the image quality is already superb on the GeForce3/4. All that nVidia needs is an increased maximum degree of anisotropy, and improved performance. Since increased max degree need not take up significantly more transistors, and improved performance should be related to the overall performance of the processor, I don't think either is much of a problem.

In other words, the only way that nVidia shouldn't have a very easy time in overcoming the R300 in is FSAA image quality. Everything else is almost a given. Of course, other things, such as anisotropic, may be merely "on par" with the R300. I suppose I should have put that in as a stipulation...that everything should be as good or superior, not necessarily superior.

 

Lets see it again.

1. In features NV30 will have more, but if it will be usefull during the next 3 years for gaming than it is another problem.

2. NV30 will use a .13 micron process, but this process show only a 30% performance advantage over a .15 micron process (TSMC data), but R300 is very well designed for such a monster chip (hand tweaked) and is using a extra power which is an advantage (if NV30 dont use extra power too).

3. Probably NV30 will use a DDR-II which is an advantage, but we are not sure if it will have a 256bits memory bus. ATI say that R300 is DDR-II capable, but no product announced using it yet.

4. Image quality I suppose both will be great.

5. The die size should be equal 120M at .13 = 96M at .15 . My guess the .15 micron will have better yields.

My guess the consumer decision will be price and R300 will have some time advantage which can translate in good prices.

edited: english

 

Well, we'll see. If it's anything like PS 1.4, then these features will be used for additional performance (mostly).

Note that it's well designed "for such a monster chip." The NV30 doesn't have many of the problems that the R300 had during design, since it's being built on a more "roomy" process, but nVidia has more engineering muscle, and a better track record. And the extra power is not an advantage. It's a necessity based on the monstrous transistor count combined with .15 microns and high clock speeds.

The real question is: What memory bus will be used with DDR2? It's very possible that neither card will be capable of 256-bit + DDR2, but will do 256-bit + DDR or 128-bit + DDR2, similar to some MX cards.

Equal? I don't think so. The R300 has quite a few more than 96 million transistors. Also, while the .15 micron process will be cheaper in the short-run, the .13 micron process will be cheaper in the long-run, especially since ATI is going to have to design two chips just because they chose to do .15 micron first.

In other words, the .15 micron R300 may be cheaper to produce at the launch of the NV30, but the NV30 will be cheaper to produce a couple of months later (And market price will depend more on demand than production costs...which means the NV30 will likely be a bit more expensive due to nVidia's proven track record and higher acceptance).

 

you fail to understand that neither company actually produces the chips. The fab company's burden the cost of production, because they have contracts with ATi and NVIDIA to produce X amount of chips... so technically, since neither ATi or NVIDIA produce their own chips, they don't get hit with fabrication size costs --- although if they wanted a contract for a chip with 200Mt's on .15, they might have to pay the premium

 

Btw, I believe I've found one thing that the NV30 could do that the R300 cannot: Procedural texture filtering. That is, if you attempt to do procedural textures on the R300, to my knowledge, you wuill get no texture filtering at all. On the NV30, you could get texture filtering by using the DDX and DDY functions included in the instruction set.