Are you ready ? (Nv 30 256 bit bus ?)

[Bjorn]

Well, nividia has some more "info" on the "are you ready" page.

This statement is maybe of a little bit of interest:

This new GPU uses the world's fastest memory technology, delivering higher performance than any other GPU-there's simply no substitute for effective memory bandwidth," he says

To me, that indicates a 256 bit bus.
Now, there's of course some "complains" about the R9700 in it also. Who would have thought ?

"Likewise, limiting pixel shader programs to just one or two hundred instructions is way too restrictive. Most of the really interesting shaders that we've seen require as many as 500 instructions, and some require even more. What we're talking about here is the ability to make games and interactive content look exactly like movie special effects do. Higher limits will allow developers to make truly realistic objects that appear to be made of real materials, with realistic reflections and shadows, all in a single pass-and all in real-time. This is simply not possible without these higher limits," he says.

Now, the obvious question is, will those 500 instruction shader run in 30+ fps ? Because he specifically mentioned games afterwards. Seems a bit unrealistic to me but who knows..

more mud throwing:

Some folks in the industry ask: "Is the games industry ready for this level of advancement?" Kirk says they are. "The movie industry has taught us that 128-bit precision and incredible levels of programmability are absolutely necessary for cinematic-quality special effects. Half measures like 96-bit precision are inadequate for many critical functions, such as texture addressing, geometric calculations like reflection and shadows, and so on," says Kirk

 

[Snap]

simply no substitute for effective memory bandwidth

If it has the highest actual bandwidth why limit yourself to saying effective?

 

[DemoCoder]

NVidia is already talking smack about ATI's 24-bit FP precision vs NVidia's 32-bit FP and about DX9 limits. Here come the endless flame wars about what's "enough" and "not enough" I personally think that while 32-bit FP is preferred, 24-bit is so much better than what we used to have that it doesn't really matter for games all that much. (offline CG rendering may be different, or CAD, etc) Not only are you going from 8-bit integer to 24-bit FP, but you also increase your dynamic range tremendously because of the exponent.


Ultimately, what it's going to come down to on these boards is: #1 which card runs most games (present games) the fastest and with good IQ at the same time. (e.g. IQ without the performance hit) The other is, #2 on a microbenchmark level, how much faster is the NV30 at executing shaders?

It may even turn out that the NV30 is slower at vertex shaders, but faster at pixel shaders.

Abstract questions like "whose pipeline is more flexible/general purpose" I think are irrelevent at this point.

I personally think that the NV30 will turn out to be "on par" with the R300, e.g. both will be great cards about the same performance, but each with their strengths and weaknesses.

Whether or not this is taken to heart, I would bet against it. Let the flamewars begin.

 

[Heathen]

Common Sense

Whether or not this is taken to heart, I would bet against it.

I'm with you, these two graphics cards represent such a huge jump in capability that for the most part their differences are only going to be of interest to the hardcore fans.

 

[nAo]

To me, that indicates a 256 bit bus.
Now, there's of course some "complains" about the R9700 in it also. Who would have thought ?

I don't think so, he says 'effective memory bandwith'.

Now, the obvious question is, will those 500 instruction shader run in 30+ fps ? Because he specifically mentioned games afterwards. Seems a bit unrealistic to me but who knows..

No way....It could be faster than r300 at pixel shaders, nevertheless it would be an order of magnitude slower of what is needed to run those very long shaders at acceptable frame rates.

ciao,
Marco

 

[demalion]

Well, nividia has some more "info" on the "are you ready" page.

This statement is maybe of a little bit of interest:

This new GPU uses the world's fastest memory technology, delivering higher performance than any other GPU-there's simply no substitute for effective memory bandwidth," he says

To me, that indicates a 256 bit bus.
Now, there's of course some "complains" about the R9700 in it also. Who would have thought ?

It actually sounds to me like the opposite, and a 128-bit bus at higher frequencies: emphasized word is "faster" with "effective memory bandwidth" as a supporting description. I'd think 256-bit would be more clearly implied by the phrasing if it were the case. Actually, I would think a 256-bit bus would be more likely in general, but this phrasing seems to me to indicate that guess would be wrong.

And to his list of what functions are necessary to be done in 128-bit instead of "96-bit" half measures...does the R300 do any of those in 128-bit, or does that "happen to be" an exact list of the things it uses 96-bit processing for? I thought it did 128-bit texture addressing for example...I guess I'll have to do some searching or review the Beyond3D article.

 

[Hellbinder]

No they dont have 256bit bus. Its 128.

Why the hell does Nvidia have to be such a bunch of slimeballs? There is no need for them to post crap lke that. Just state what your card does and STFU. Making flippping low-life slams on Ati and in the samne sentence printing laughable crap liek this.

GPU-there's simply no substitute for effective memory bandwidth

No, David Kirk you freaking Parolle. *REAL* bandwith is better than *effective* bandwidth. You guys have 16GB of raw bandwidth and you damn well know it. So stop the BS.

It frikking kills me that you guys all laugh about the flipping total slimmy BS Nvidia does like its cute. I dont remember ATi posting derogatory **** about the GF4 in their press releases. All they did is talk about their own Tech. How can you people like a company that conducts such underhanded Trash tactics. It makes me ill.

 

[Dave Baumann]

I dont remember ATi posting derogatory **** about the GF4 in their press releases. All they did is talk about their own Tech.

You must have a mental block on their whole bunch of PR slides that effectively rubbished the GF4 line whilst they launched their 128MB boards then. Odd, seeing as it was mentioned only a couple of days ago.

 

[Hellbinder]

The only one i remember was the slide regarding the naming scheme...

i suppose you could be right....

 

[nAo]

]No, David Kirk you freaking Parolle. *REAL* bandwith is better than *effective* bandwidth. You guys have 16GB of raw bandwidth and you damn well know it. So stop the BS.

It depends. What does 'effective' mean in this case? We don't know.
Is it just a raw bandwith figure multiplied by compression ratios or is he addressing the 'effective' bandwith the hw
can exploit out of the theoretical bandwith? What if nv30 spanks on a plurality of benchmarks the r300 with 20 % or 30% less bandwith?
Wouldn't you believe a 4 way crossbar controller on a 256 bit bus could be thoeretically as efficient as a 4 way crossbar on a 128 bit bus with the latter having better granularity?
When first numbers will hit the net you (and we) will be free to do a <final> judgement, but at the moment you can't.

All they did is talk about their own Tech. How can you people like a company that conducts such underhanded Trash tactics. It makes me ill.

Well..I don't have the right links to post at the moment, but believe me, you need a reality check on this.

ciao,
Marco

 

[Entropy]

Some folks in the industry ask: "Is the games industry ready for this level of advancement?" Kirk says they are. "The movie industry has taught us that 128-bit precision and incredible levels of programmability are absolutely necessary for cinematic-quality special effects. Half measures like 96-bit precision are inadequate for many critical functions, such as texture addressing, geometric calculations like reflection and shadows, and so on," says Kirk

lol.
I'd like to see where 24-bit/component FP is insufficient, whereas 32-bits would be OK. Frankly I'd like to see any non-amplifying graphics algorithm with a gaussian error distribution where the errors from using 24-bits per component exceeds the limitations of the output medium. 16-bit FP is marginal. But 24-bits? Hmm.

Plus, it isn't that long ago that Kirk claimed that > 8 bits per component was overkill. Regardless of pedigree, the man is too clumsily obvious when he plays the role of marketeer. He always points to the areas where nVidia is doing well as absolutely critical, and the areas where they are behind are not only neglected, but (and this is where he is overdoing it) he actively claims that they are insignificant. Plus, he misrepresents the opposition.

Leaves the sour taste of bad politician in your mouth.

Entropy

 

[demalion]

NVidia is already talking smack about ATI's 24-bit FP precision vs NVidia's 32-bit FP and about DX9 limits. Here come the endless flame wars about what's "enough" and "not enough" I personally think that while 32-bit FP is preferred, 24-bit is so much better than what we used to have that it doesn't really matter for games all that much. (offline CG rendering may be different, or CAD, etc) Not only are you going from 8-bit integer to 24-bit FP, but you also increase your dynamic range tremendously because of the exponent.

Well, if they can provide an image that actually looks different to me based on the difference in precision, I say let them talk "smack" . I don't think they can, but they are welcome to prove me wrong.

Ultimately, what it's going to come down to on these boards is: #1 which card runs most games (present games) the fastest and with good IQ at the same time. (e.g. IQ without the performance hit) The other is, #2 on a microbenchmark level, how much faster is the NV30 at executing shaders?

Well, number one has always been the criteria before, hasn't it? And #2 should be part of #1 next year around this time, I'd think.

It may even turn out that the NV30 is slower at vertex shaders, but faster at pixel shaders.

Abstract questions like "whose pipeline is more flexible/general purpose" I think are irrelevent at this point.

Yes, they are, to gaming. But we've been discussing that already and we'll likely continue to discuss it. We've managed for the most part not to turn it into a flame war, haven't we? It's when the old "phallic fps" syndrome comes into play that the flaming starts I think. Also, with a continued commitment to mud slinging, I think nVidia marketing (which Kirk seems to be pretty solidly part of now ) is going to provide the source for most of them. It will be interesting to see whether ATi tries to "take the high road" in responding them (addresses the comments and provides info and counter arguments), stoops to the same level (makes disparaging comments that can be clearly seen to refer to competitors and relies on unsupported implication), or tries to ignore the comments.

EDIT: BTW, Hellbinder, ATi has directly attacked GF 4 naming, as DB pointed out. What they haven't done (yet), to my knowledge, is offered mudslinging in the guise of "technical information" (sometimes even wrong technical information) about a competitor. David Kirk seems to make that his job...would anyone disagree with this? I'd sort of hoped that nVidia might be outgrowing that...I suppose I'd better take a full look at this text to see.

I personally think that the NV30 will turn out to be "on par" with the R300, e.g. both will be great cards about the same performance, but each with their strengths and weaknesses.

Whether or not this is taken to heart, I would bet against it. Let the flamewars begin.

Well, there is more fuel for the fire for differences because the featuresets are even more complex and extensive, and more opportunities for mud slinging (witness the comments we are discussing ). The only thing that will discourage flamewars is the inability to see a difference in produced images. That, and both cards are likely to be simply "overkill" for atleast the time period until Doom III ships (maybe Deus Ex 2 or Thief 3 with their enhanced shadowing might qualify as well if they are released earlier), and upon the release of these games we'll be back to the fps obsession I'd think.

But it is quite a noble sentiment.

 

[Humus]

Now, the obvious question is, will those 500 instruction shader run in 30+ fps ? Because he specifically mentioned games afterwards. Seems a bit unrealistic to me but who knows..

I'd be very surprised if such a shader would run at 30+ fps in anything higher than 160x120. My latest demo has a 26 instructions in the fragment shader, and that's enough for getting a R9700 down to below 60fps in 1024x768. Given such stats, linear scaling and assuming the NV30 is twice as fast as a R9700 it would run at 15fps at 640x480.

 

[andypski]

It's the mixed messages in market-speak statements that I find amusing. They change diametrically depending on the circumstances

Marketing 101
- If you don't have a comparable product, rubbish your competition until you do.

There are apparently no effects you can produce with PS1.4 that can't also be produced with PS1.3 - which has less instructions per program, and is less flexible. This is a pretty safe statement since you can do pretty much anything if you multipass enough times...

Of course multipassing an effect with a pixel pipeline that has, let's say, the rather poor 'half-measure' of only 9 bits of total internal precision vs. a single pass with a pipeline with rather more bits of internal precision will quite possibly get you a quantised mess, but at least the mathematics are the same even if the results aren't...

Now we move forward to today. Apparently now 100+ instructions (roughly 10x the last generation) just isn't enough, and really you need more than 500 to do anything really interesting. Multipassing (with intermediate buffers and pixel pipelines of much higher precision than were ever available in the past) miraculously doesn't enable you to reproduce these interesting effects...

The 'revolutionary' per-pixel shading effects of the 'NfiniteFX (TM)' 1.3 shaders weren't really that interesting at all... I guess they were 'finiteFX' after all...

The new market-speak has us talking about filling screen pixels with a 500+ instruction shader (roughly 50x more complex than any 1.3 shader can be.)

A quick and dirty analysis of how fast a GF4 can execute a full-length shader program will tell the interested reader how much more powerful a new product would need to be to run these 'interesting' shaders at an interactive frame rate...

(Quick hint - it's somewhere between a sh*tload and a f*&kload, if you'll pardon my colourful metaphors).

And, we won't mention that the longer a program becomes the smaller the relative performance hit from multipassing becomes...

 

[Humus]

Heh, I wonder how nVidia will market that the NV30 operates at twice the speed in 64bit mode than 128bit mode given than 96bit is "inadequate".

 

[Humus]

Saw this at HardOCP:

 

[Joe DeFuria]

"The smaller 0.13 micron process allows us to put more transistors in each square millimeter than we could achieve with 0.15 micron," says Kirk. "Also, due to the fact that each transistor is smaller, these transistors are also faster, and each one uses less power." How great is that? "Smaller and faster-no compromises."

Well...no comprimises except for missing a product cycle.

In any case the main difference bewteen this, and ATI's previous campaign is not the mud-throwing itself. I believe ATI's "campaign slogan" for the 9700 launch was "crush the Force" or something to that effect.

However, if yo urecall, ATI's other slogan was "no paper tigers...let the product speak for itself." Now, while ATI DID speak for the product, they also let the product speak by:

1) Benchmark comparisons using actual hardware
2) Visual and direct comparisons (renderings) of what 9700's pipeline offered vs. al other existing technologies.

Now, this little "pre-launch hype" from nVidia is a bit distatseful, but it's not really much to be concerned about. What should concern everyone is the official launch. If all nVidia has is in fact "paper tigers", that should immediately throw up some flags.

 

[LeStoffer]

I personally think that the NV30 will turn out to be "on par" with the R300, e.g. both will be great cards about the same performance, but each with their strengths and weaknesses.

I have the same 'feeling', but what was really caught my attention is the fact that nVidia basicly just restate what we already have known about NV30 strenghts [over the R300 line] from the CineFX papers. And for a long time I might add.

It would be about high time that they started to hype other strenghts of the NV30 - if there is any at all besides increased clockspeed. And this leads to my point: More advanced shader abilities and higher clockspeed might well be all they have to attack the R300 line with.

 

[BRiT]

I do believe the NV30 will not have a true 256-bit bus. Everything that NVIDIA has been saying is effective and smarter. As for the need for over 500+ instruction shader programs, I doubt that. There hardly exists any games today and in the next 9-12 months that will make use of DX9+ shaders. Yes, all those extra features are nice, but will sit mostly useless with exception to demos and synthetic benchmarks.

If ATI is aggressive, it wont be the R300 lines that NVIDIA/NV30 will be attacking, it'll be the R350 lines.

--|BRiT|

 

[antlers]

How fast might the NV30 be able to run even a 100-instruction pixel shader, even at 64 bits? Might they pull a miraculous instructions/clock figure out of their hats, after all?

Given the experience of DX8, won't it be like 2 years before any games that feature DX9 pixel shaders are available? Given that DX9 is so far beyond what games make use of today, is it likely that any developers will target capabilities beyond DX9?