NV30 = no 256 bits wide data bus?

[nAo]

I just read this interview to David Kirk. The interviewer ask to him if nvidia is going to use a 256 bit data bus on the NV30 architecture. Mr. Kirk answers: (very rough translation)

I can't share any technical details on NV30 memory architecture, but I would answer to you pointing out that a wider data bus gives more bandwith, but with a cost: a 256 bits wide architecture is very complex and expensive too. High-end boards are very expensive. Ask to customers to pay a higher price to integrate a 256 bits wide bus it's an overkill at the moment.

ciao,
Marco

 

[Natoma]

If NV30 is in fact the amalgamation of Fear and whatever Nvidia had on the drawing board at the time, then they won't need a 256bit data path.

They can get along just fine with 128 bits.

 

[John Reynolds]

Fear was a Gigapixel-based tiler, and how many times have we heard from Nvidia that they simply won't go that route?

 

[Tagrineth]

Bah!

They throttle us with the $400 Ti4600 then turn around with this and say they don't want to overcharge

 

[Mintmaster]

Bah!

They throttle us with the $400 Ti4600 then turn around with this and say they don't want to overcharge

How about the $500 GF3 (Actually $600 on the Mac first), and GF2 Ultra?

 

[Joe DeFuria]

Hmmmm....

Based on the "babelfish" translation of the entire response to the "256 bit" question, it sounds to me that Kirk is setting up NV30 to not be "faster" than current generation products, but have "higher quality pixels."

Here is my "translation" of the babelfish translation.

The true problem, in my opinion, is to determin if such technology (256 bit bus) today is demanded and needed. I'll explain myself: if in order to improve the quality of the experience 3D, it's not necessary to generate more pixels, but it's more important to render every single pixel more accurately (not more pixels, but better pixels), need of more and more complex and fast GPU, not increased video memory bandwidth. If we don't increase the number of pixels on the screen, in fact, a increased bandwidth does not allow us to obtain a qualitative improvement of the generated image; it would help in presentational terms if they used more pixels, but at the moment with the taken resolution of 1600x1200 @ 32bits as parameter of reference for the games 3D we do not see the need of having more memory bandwidth, regarding how much today we can supply with the card GeForce 4 Ti4600.

Sounds to me like NV30 won't have any more bandwidth than GeForce4. I do agree that beyond 1600x1200x32 there is diminishing returns as far as raw resolution goes. But "higher quality" pixels to me means higher color precision, better filtering, and better AA...all of which require more bandwidth. So unless NV30 has some techniques to do something like 1600x1200x32 with high anistotropic filtering ans 16X edge AA at 60+ FPS with the same bandwidth as GeForce4...then Kirk is just blowing smoke.

I'm betting that nVidia is going to go their usual route...that the "higher quality" pixels are obtained with DX9 programming. NV30 won't be much faster than GeForce4 at DX7/8 apps, but we'll "need" one for a DX9 x-mas.

 

[Joe DeFuria]

A few more of my "translations" of babelfish translations, concerning NV30 Q&A:

Q. At least one of your competitors will release a solution with 256 bit memory bus, therefore raising the bar.

A. Probably this compeitor has opted for one solution that is contrary to our opinion...they are going to supply more pixels as a means of delivering higher quality. It's a matter of different approaches, and I believe ours is better.

Q. Will it be possible to have DDR Memory with frequencies much higher in the next few months? The near-term limits seem to be between 700 and 750 Mhz, with the most recent chip from Samsung at 2 ns that they still have yet to debut on the market.

A. The next generations of cards will come with faster memories. The problem, however, is not just one of having more bandwidth, it's how well you take advantage of it.

Q. Hardware Displacement Mapping in DirectX 9: what you can say to us?

A. ....obviously in nVidia we know many things about the new API but, having signed a NDA with Microsoft, I cannot comment.

 

[siconik]

If Nvidia thinks this will fly, they are horribly mistaken. Have they learned nothing from GTS VS V5? Consumers are much more interested in the "faster" pixels then in "better" pixels. Besides, the only way to deliver these "better pixels" is by having plenty of power (bandwith/fillrate) to spare.

 

[ChrisK]

...

Heh, could Nvidia be caught on the wrong foot this year?

If they're going down the "not much faster but many additional features" then their life could be somewhat miserable from next week on when Matrox introduces their new part...

Would be funny to see them desperately scrambling for more performance

 

[Livecoma]

As far as being competitive goes I don't think It's NVIDIA we should be questioning. I have faith in their ability to go a step above the competition, and they have proved that more then once in the past...

I'm more intrested in how certain other companies plan to compete this fall as I think they have more to proove...

 

[CoolAsAMoose]

If Nvidia thinks this will fly, they are horribly mistaken. Have they learned nothing from GTS VS V5? Consumers are much more interested in the "faster" pixels then in "better" pixels. Besides, the only way to deliver these "better pixels" is by having plenty of power (bandwith/fillrate) to spare.

Consumers may NOT be "much more interested in the faster pixels then in better pixels", but with Q3 still being the most important benchmark, and the most important test being without AA, I agree that this may be a big mistake from NVidia. And besides the Q3 numbers, for OEM sells, numbers are more important than features: "256 bit memory bus" sells better than "Supersmooth pixels".

 

[Bjorn]

... but with Q3 still being the most important benchmark, and the most important test being without AA, I agree that this may be a big mistake from NVidia. And besides the Q3 numbers, for OEM sells, numbers are more important than features: "256 bit memory bus" sells better than "Supersmooth pixels".

I don't think that Quake3 will be such an important benchmark this coming winter. We already have the Unreal performance test and maybe there'll be a Doom3 test also before the NV30 arrives.

And i don't think poeple will prefer a 300 fps Quake3 and 30 fps Doom3 (4X FSAA and anistotropic. I don't really buy that the "without FSAA" still are the most important test) cards rather then a 200 fps Quake3 and 100 fps Doom3.

Now of course, we don't really know if the NV30 will be any better in f.e Doom3 then the competition. This winter will probably be one of the most interestings times for us 3D freaks. A new Unreal engine, new ID Engine and some really interesting new 3D cards. This time maybe from 4-5 vendors instead of just 2. The NV30 will certainly face some tough competition.

Summary:

Quake3 important for the NV30 + other next generation cards ?

I certainly don't think so.

 

[kl899]

Summary:

Quake3 important for the NV30 + other next generation cards ?

I certainly don't think so.

Don't forget that none of us on this board are average consumers. We have highly opinionated views of 3D performance. I think the answer is that the average consumer DOES care about raw speed in existing games. That's what they can brag to their friends about, etc. Let me tell you something... my roomate went from a GeForce 2 MX to a GeForce 3 TI 500. Guess why he made that switch? To make his Counter Strike framerate faster. Both of us know that the GeForce 2 MX is plenty enough to run Counter Strike but people always want to have the fastest hardware, not necessarily for more advanced quality features.

 

[Joe DeFuria]

I don't think that Quake3 will be such an important benchmark this coming winter. We already have the Unreal performance test and maybe there'll be a Doom3 test also before the NV30 arrives.

I agree that this fall / winter, Quake3 will have a lesser role in benchmarking. However, I don't think it will go away completely, as it will tend to be the "representative of past game / fill-rate limiting" performance benchmark. Just like we have seen Quake2 and Quake3 benchmarks in parallel for a while, the same will be true of Quake3 and UT Test / Doom3.

As far as FSAA goes....this is where things are going to get hairy. The tendency, of course, is for the cookie-cutter sites to just run bencharks with "2X FSAA" enabled, and comparing FPS...nevermind that each vendor probably has their own way of doing FSAA and trying to compare actual image quality between the two. So I agree that FSAA (and advanced filtering) tests SHOULD be more important this fall, I have low expectations of review sites actually doing it "right". And the "winner" will be the board that simply gets the higest FPS at the equivalent "setting". (Again, setting meaning "2X, 4X" etc., not necessarily equivalent image quality as it should.)

Now of course, we don't really know if the NV30 will be any better in f.e Doom3 then the competition. ...

and the thing with Doom3 and UT Test...is that they are not DX9 type apps. They are actually DX7 apps at heart, but have some DX8 functionality. So if the only "advantage" of NV30 (or any next-gen card) is "merely" that it has DX9 functionality, and won't typically do DX7/8 apps any faster...then even Doom3 and UT Test might not show any performance advantage on the latest hardware.

Enter 3DMark 2002/3...

 

[Ascended Saiyan]

If I'm not mistaken the problem that we have in today's graphic cards are either having not enough bandwith or just not utilizing the bandwith properly.Nv-30 will be the first supposed chip from nvidia with 3dfx tech & Nvidia since seems to have ruled out both Tiling as well as a 256 Bit memory interface & this can mean only two things either they have found an extremely conservative way of handling bandwith or they will be jumping on the Bitboys bandwagon & utilize E-dram because to me this can be the only possible way that they can use most of the new Direct x 9 features especially the most bandwith intensive being the enhanced color precision.

 

[Joe DeFuria]

this can mean only two things either they have found an extremely conservative way of handling bandwith or they will be jumping on the Bitboys bandwagon & utilize E-dram ...

I personally doubt EDRAM.

As for just handling bandwidth more efficiently without "radical" changes, I see this as being done a few ways:

1)Continued evolutionary changes in memory interface, z compression, etc. Self explanatory.

2) Re-orgainization of pixel caches for AA / advanced filtering. I have read at least one vendor make a comment like there being a "shift" from optimizing pipelines from "non-AA'd" pixels, to AAd pixels. In other words, be re-orienting the pipelines / cache to focus on delivering some level of AA as the baseline, I can see good gains in efficiency when AA is enabled. The "drawback" being that when AA is turned off, you won't get as good performance as you would have if the pipelines were designed as they traditionally have been. I believe the time is about right to start to take this approach. However, the risk is having such an architecture being run "without" AA enabled. If nVidia or someone else takes this risk, you will see HEAVY evanglization for "all benchmarks should now be run with AA enabled."

3) Increasing pipeline flexibility. I still don't really have an idea if the DX9 pipeline requires more or less "bandwidth" than the DX8 one. Though, I would think (someone correct me if I'm wrong) that as long as you don't have to resort to multiple pass rending with DX9, you should be able render a "higher quality pixel" with less bandwidth than trying to render the same pixel with DX7 or DX8. So in that sense, bandwidth is being used "more efficiently." In reality though, developers just don't render DX9 pixels in DX7/8. So this means the efficiency is only actually realized when someone codes specifically for DX9.

 

[pascal]

My guess for NV30:

- 128bits 400MHz DDR (12.8 GB/s)
- Same memory interface LMA II.
- Very large/better Quadcache. Specially large and better associativity textures caches.
- Better oclusion culling.
- Single pass up to 8 level multitexture.
- DX9 features

Conservative evolution.
Then it could be twice as fast as GF3 when playing Doom3, and it means 60fps. The question is how much will it cost and what about the competition?

3Labs is not big competition for this year (Tom´s hardware say the price will be $890 for 128MB cards and $590 for 64MB). What about Matrox and ATI? Any new player?

 

[psurge]

Can the nv30 afford very large caches when it already must use FP throughout the pipeline?

Question. Since DX9 appears to support 1,2, and 4 channel 32bit floating point formats for the frame-buffer, won't the cost of a single pixel shader unit (minus texture lookups) be similar to the cost of a single vertex shader (somewhat less since there are less constant registers/instructions) transistor wise?

For some things I can see fixed function hardware always being around : i.e. clipping, culling, setup, HZ-buffer, texture loads/filtering.

But for the rest, why even have separate types of functional unit?

Given the fp frame-buffer requirements... why would you implement VS2.0 and PS2.0 as different kinds of functional units? Wouldn't it make sense to have one unit that can do it all, optimize it as much as you can, and then put as many of those on the chip as you can fit?

Serge

 

[Joe DeFuria]

3Labs is not big competition for this year (Tom´s hardware say the price will be $890 for 128MB cards and $590 for 64MB). What about Matrox and ATI? Any new player?

Actually, those are for the "professional" versions of 3D/Creative Labs' P10 boards which are due out in the next month or two. We have no idea what Creative will charge for the "consumer" version of the P10, which is due out "by X-Mas". Certainly, it doubt it will be as high as $600 for a 128 Mb board.

What we don't know, is how technically, the "consumer" version will differ from the professional version. It may have a cut down core, cut down bandwidth...it might be a die shrink. It might be nothing more than the difference between the Quadro and GeForce boards...which is "physically" about nothing. Just some tweaked drivers and artificially "turned off" features on the consumer version.

I'll be buying a new card this fall / early winter. And I fully expect to have to choose between nVidia NV30, ATI R300, 3D Labs P10, and Matrox Parhelia. Can't wait.

 

[SA]

Prices are pretty much determined by the market and competition.

Creative/3DLabs should be trying to gain market share in the consumer market over the next couple of years, so you can expect them to price their products accordingly.