NVIDIA's take on GDDR-3

ben6

Regular
We are big fans of faster graphic memory and can use all the speed memory makers can provide. The GDDR-3 spec is a work in progress. Generally it is a good spec and we are glad ATI is evangelizing graphics memory needs with memory vendors, which is something NVIDIA has been doing for years. We think the spec needs to be more aggressive. NVIDIA cannot endorse the spec as is. It needs more work. 500MHz as a starting point is boring in the time frame they are talking about, which is 2003. 500MHz in 2002 is exciting. 500MHz in 2H03 is boring. The value of the GDDR3 as it stands in the current spec will be severely tested in the coming weeks with existing technology.
Brian Burke
NVIDIA Corporation
 
Interesting. That more or less confirms that NV30 (at least one version of it) will ship with 500 Mhz DDRII.

To me, that also imples that NV30 will be a 128 bit bus. I just don't see NV30 shipping with both exotic memory AND exotic memory bus, while still being somewhat affordable at this time.

That being said, I would love to see 500 MHz DDR-II on a 256 bit bus!

If 128 bit is true, then of course ATI's answer would be "well, 128 bit bus in a product shipping in volume in 2003 isn't exciting...256 bit bus in 2002...now that's exciting.

Gotta love PR! ;)
 
Regardless of actual memory bandwidth, it's the end performance that counts (particularly with FSAA/aniso).

Personally, I don't see any problem mixing high-speed memory with a 256-bit bus. And as for "somewhat affordable," the NV30 is obviously going to be a high-end product.

Still, I do hope that nVidia manages to ship with a 128-bit bus that outperforms ATI's R300 (with FSAA/aniso, of course). That would make for one amazing video card, however unlikely it may seem.

As a side note, to fill a 256-bit memory bus at 500MHz memory, nVidia would need to either run their core at 500MHz, or have a huge on-die memory controller (i.e. larger than 512-bit).
 
Heh, I just got this response from Chris Hook (who I met at ATI's 9700 launch in San Francisco nice guy):Hi Ben!!
Saw your article on GDDR-3 and Brian Burke's comments. It sounds to me like there is some disagreement at nVidia over the GDDR3 issue and whether or not the GDDR3 spec is 'fast enough'. I would point you to some comments by Tony Tamasi in EBN Magazine a few weeks ago, when ATI's GDDR3 spec was announced, with the support of DRAM memory industry leaders. Mr. Tamasi was quoted as saying that GDDR-3's very high speed will definately be welcomed by graphics companies. See link and quote below

Chris Hook
PR Manager, Mobile and Integrated Business Unit
ATI Technologies

http://www.ebnonline.com/showArticle.jhtml?articleID=3000038

Nvidia said that it, too, is working with DRAM suppliers to get GDDR-III to market as soon as possible. "If we can get enough DRAM makers to build GDDR-III, it should become very affordable," said Tony Tamasi, Nvidia's senior director of desktop product management. "The very high speed will definitely be welcomed by graphics companies."
 
Joe,

Albeit one can never have enough, do we really need at this point 32GB/s bandwidth?

As far as PR goes, send both greetings and tell ´em that the only thing that is exiting is efficiency within an architecture.
 
Bandwidth is one of those things that you can never really have enough of :LOL: . Memory bandwith will continue to play an important role in the 3D world. Efficiency is also extremely important, and if can have a architecture that is efficient without being dependent on render-order and thus on the games developer, it`s all for the better.Granted, if you have a greatly inneficient architecture like Parhelia, not even heaps of bandwidth will save you.
 
Granted, if you have a greatly inneficient architecture like Parhelia, not even heaps of bandwidth will save you.

That´s what I meant more or less.
 
Personally, I don't see any problem mixing high-speed memory with a 256-bit bus. And as for "somewhat affordable," the NV30 is obviously going to be a high-end product.

I don't see any "problem" except very high cost. I said "somewhat" affordable because at $400 for the typical high end card release, that's not really affordable at all. If nVidia sells a 256 bit board equipped with 128 MB of 500 Mhz DDRII at $400, then ATI must be really taking in the profit margins on the Radeon 9700....

And for nVidia's sake, there had better be more than one "high priced" $400 NV30 variant at launch...because having only that one product against ATI's line-up won't do much good.
 
Joe DeFuria said:
And for nVidia's sake, there had better be more than one "high priced" $400 NV30 variant at launch...because having only that one product against ATI's line-up won't do much good.

A. nVidia's still got more credibility in the marketspace, so they're not going to have to try overly-hard to push a good product.

B. Whether or not multiple variants will be available will be determined by how many chips they can produce. There's a good reason why nVidia has traditionally only released one variant of every new architecture for the last few years right at launch (The GeForce 256 almost breaks this, but the DDR wasn't available widely for a little bit after launch...).
 
That being said, I would love to see 500 MHz DDR-II on a 256 bit bus!

That's exactly what I want to see as well!

If 128 bit is true, then of course ATI's answer would be "well, 128 bit bus in a product shipping in volume in 2003 isn't exciting...256 bit bus in 2002...now that's exciting.

well said! that would be my counter also.

I don't see why Nvidia can't have highend memory combined with 256-bit bus. All of the other new graphics processors (R300, P10, Parhelia) have a 256-bit bus, why not NV30 as well? I hope it does.

here's what I want to see from Nvidia:
NV30 - 8 pipes - DDR-II - 256-bit bus - upto 256MB 'highend'
NV31 - 4 pipes - DDR-I - 128-bit bus - 128 MB 'midrange/mainstream'
 
I found this report interesting. However, if everone has their own "graphics DRAM specification" under the table, wouldn't that makes memory cost higher?
 
Megadrive1988 said:
I don't see why Nvidia can't have highend memory combined with 256-bit bus. All of the other new graphics processors (R300, P10, Parhelia) have a 256-bit bus, why not NV30 as well? I hope it does.

hopes like these, usually generates products with extremely high price. and that's something that I don't like. nVidia has been usually record breaker in performance as well as in prices. (at least here in northern europe.) So, I think they will do it this time again, even without combination of exremely fast & expensive memories and expensive board implementation. (anyone has any idea what kind of PCB is needed for 256Bit bus running on extreme speeds like 500MHz DDR?? In my logic, when the clocks go higher and you have more inducts, you need more layers to get same signal/noise ratio than with thinner bus and/or slower clocks.)
 
I think 128 bits is much more manageable (i.e. more efficient usage) as well. Look at the 9700 vs. the 9500, which both have the same memory and core clock speeds. In the single texture fillrate test:

9500: 969 Mpixels/s
9700: 1532 Mpixels/s

You can see that double the theoretical bandwidth only gave the 9700 a 60% advantage, and it's bandwidth limited on both. 500 MHz on a 128 bit bus should probably have a bit more effective bandwidth than the 9700PRO due to inefficiencies in ATI's 256-bit bus, all other things being equal (which they likely wouldn't be - who knows what other crazy stuff NV30 will have).

Well, in any case, NVidia knows what they're doing. Maybe they won't even have 8 pipes, but 4 that can do many ops/textures per cycle, and a crapload (32 or 64?) of Z/stencil-check pipes for both MSAA and hidden pixel rejection.

Whatever, I'm just guessing. But 128 bit seems very real, and not as stupid as some people make it out to be.
 
well aren't they talking a 1Ghz ram ? I would think that 1Ghz ram by its self would be very very expensive and then to add a 256 bit bus might well break the bank big big time
 
A. nVidia's still got more credibility in the marketspace, so they're not going to have to try overly-hard to push a good product.

Hmmm...so far, it looks to me like they are already pushing very hard, before the product is even officially announced! In any case, "credibility" is only really as good as your current product product. 3dfx once had lots of credibility too.

OEMs are just as fickle as consumers. "Credibilty" can turn almost a dime. They don't care abobut tech, they care about selling products. I would certainly agree that "one late product" doesn't ruin the credibility of nVidia...but I would also argue that if ATI bests nVidia for two product cycles in a row, "credibility" will rapidly shift. This spring is going to be VERY important for nVidia. If ATI does have answers for whatever nVidia is cooking up...then that puts nVidia in a very, very unenviable position.

There's a good reason why nVidia has traditionally only released one variant of every new architecture for the last few years right at launch...

Yes, and that was becase nVidia could get away with it, because the competition had no "answers" to that one variant.

This situation is very different now. ATI has arguably the best product in every market segment (shipped or announced). nVidia is basically no longer "competing against themselves" as I believe you used to say quite often. When you are in that position, you can certainly afford to launch just one or two high end variants, and let your "older" parts stay in the mainstream.

There's also a good reason why nVidia switched from pushing the MX 460 to the Ti4200 as the "Sub $200" card. Competition forced them to ship a product with lower margins. Competition is really "forcing" nVidia to ship something better than AGP 8X variants of the GeForce4 MX and Ti-4200. The question is, does nVidia have such a product in the pipeline due out soon?
 
Back
Top