Latest NV30 Info

[EasyRaider]

I think one of the things that came out of the NV30 / R300 article is that where R300 only colour compresses with FSAA NV30 sounds as though it may use colour compression with or without FSAA.

That sounds like a very good possibility...

How? With MSAA there is plenty of redundant data to compress (subsamples with same colour).

A lossy algorithm is a possibility, but I'm not sure if I would want that.

 

[DemoCoder]

I can't fathom an NDA'ed source who likes posting information that may be uninteresting or worse, even damaging to NVDA (e.g. saying it has a 128-bit bus without putting it into context wrt the rest of the product), but is afraid to post something relatively innoculous (but verifiable) like "NV30's new efficient architecture is called Lightspeed Memory 3, or Tachyonic Memory Bus(tm) or whatever"


I mean, if I was under NDA, what would I be more afraid of releasing info about, a new minor specific feature or trademark name, OR, saying "NV30 features are same as GF4" Both are a violation of NDA and equally culpable, but one of them is a bit of negative FUD, the other is positive FUD.

IMHO, if you claim you are a so-called insider, and release some supposedly NDA'ed info about an upcoming chipset, but when further pressed for any clarification, retreat back to the excuse "But I'm under NDA and can't say anything!", you are either a liar or a hypocrite.


If a source can't produce any information to differentiate himself from the background noise of rumors, he isn't a source.

 

[noko]

Two weeks time and we should have our answer is the real answer to this guess work.

 

[overclocked]

Maybe they skipped the fixed T-L unit to free some transistors and maybe they have moved to eight 16-bit memory controllers?

 

[Ailuros]

I personally am not aware of any insiders from any IHV that throw out information that's under NDA to individuals they're not 100% sure that it won't go public.

Of course can someone throw a couple of harmless masked as personal estimates/speculations, but speclists?

 

[OpenGL guy]

maybe they have moved to eight 16-bit memory controllers?

Why would you want 16-bit memory controllers? The majority of the data is better represented as multiples of 32-bit values. It would take two cycles to write out a single 32-bit color on a 16-bit memory controller.

 

[overclocked]

Ok, now i dont need to ask you why..

 

[Hellbinder]

Effective Bandwidth taking into account the various bandwidth saving techniques, according to some rumours, is more than twice higher

100$ says that the effective number (that they make up) will be on the box...

BTW, Radeon 9700 has really good HSR and Compression techniques as well... Perhaps ATI should claim 2x effective bandwidth as well. As for the comment i read that perhaps Nv30 uses Color compression all the time.. It seems to me that the real benefit of Color compression is in FSAA where CPU is not the bottleneck.

 

[arjan de lumens]

maybe they have moved to eight 16-bit memory controllers?

Why would you want 16-bit memory controllers? The majority of the data is better represented as multiples of 32-bit values. It would take two cycles to write out a single 32-bit color on a 16-bit memory controller.

DDR-I has a minimum burst length of 2, so with 16-bit controllers, you are still forced to access data in at least 32-bits-wide chunks (and that would be 64 bits with DDR-II).

 

[overclocked]

If i remember right Geforce4 has four 32-bit mcontrollers and Radeon9700 has four 64-bit.. Would it make any sense for ATI to go to eight 32-bit then?

Also could more cache help the bandwith significantly?

 

[Testiculus Giganticus]

I think people should be suspicious about any architecture specific info posted until the NDA is lifted.NDA infringement is not a thing to be played with. So, clock rates, number of pipes and bus width should be taken with a big grain of salt until nVidia says something official.Just my advice 8)

 

[DegustatoR]

noko

Now how would a Russian site get inside information from a California based company??

Same way as every other site on the World Wide Web.

martrox

Ok, guys, IF the NV30 has similar FSAA & AF, at similar , or just a bit quicker speeds

I didn't say this! Same algorithm, as in R100 & R200, used for aniso in R300 (w/some tweaks), but look at the difference! Same method, as in NV2x, used for FSAA in R300, but FSAA 6x on R9700Pro is clearly superior (in both speed & quality) to FSAA 4x (and even 4xS) on GF3/4.

So, NV30 may have same algorithms for aniso & AA, as were in NV2x, but these algorithms may produce better image quality with smaller perfomance hit. This is especialy true for aniso.

Username

a "reliable source" who is cobbling together information based admittedly on rumors

Again, I didn't say that! Information (the one without question marks) is correct and it's not gathered from WWW-Rumors Corp. simply because there are sooo many conflicting rumors about NV30 today (8x2 or 4x4? 128bit or 256bit? 700MHz DDR or 1GHz DDR2?), that a person who doesn't know for sure couldn't make such "compilation" without doing some really stupid mistakes.

DemoCoder

I can't fathom an NDA'ed source who likes posting information that may be uninteresting or worse, even damaging to NVDA (e.g. saying it has a 128-bit bus without putting it into context wrt the rest of the product), but is afraid to post something relatively innoculous (but verifiable) like "NV30's new efficient architecture is called Lightspeed Memory 3, or Tachyonic Memory Bus(tm) or whatever"

_I_ post what _I_ know. _I_ know what i'm interested in. I'm _not interested_ in marketing bs. So i don't know it.

And again: i'm_not_ under NDA w/NV.

you are either a liar or a hypocrite

You gotta relax sometimes, you know? Try to listen to some Ninja Tune music

If a source can't produce any information to differentiate himself from the background noise of rumors, he isn't a source.

Right. But to understand that this information is correct you a) gotta be under NDA yourself OR you b) can have some friends, who are under NDA. In a) case you will be dissapointed, that someone without NDA disclosed something, that you have no right to disclose (and you will call him "a liar or a hypocrite" ). In case b) you will simply understand that this information is correct. But in any of these cases you will understand that this source is telling the truth. And, judging by your reaction, you already understood this

There is case c) however. In this case you know _nothing_ about NV30 and simply can't judge me for "releasing" this info (remember, all this information was already leaked, when i posted it).

overclocked

Maybe they skipped the fixed T-L unit to free some transistors

AFAIK, there are no dedicated HWT&L unit in NV2x architecture. It is already emulated by special vertex shader.

Ailuros

I personally am not aware of any insiders from any IHV that throw out information that's under NDA to individuals they're not 100% sure that it won't go public.

All this info already was public in form of rumors. And these "speclists" produced by me, not by any IHV or NDAed person. You could understand this, if you'd see my original posting (http://forum.ixbt.com/0010/036907-118.html#3254, it is all in Russian, however). There were two nearly identical "speclists" for R300 (R9700Pro) and NV30, merely for better perception by readers and for clear comparison of two.

[offtopic]That was my yearly dose of English... [/offtopic]

 

[T2k]

I think it's just ridiculous...

 

[OpenGL guy]

maybe they have moved to eight 16-bit memory controllers?

Why would you want 16-bit memory controllers? The majority of the data is better represented as multiples of 32-bit values. It would take two cycles to write out a single 32-bit color on a 16-bit memory controller.

DDR-I has a minimum burst length of 2, so with 16-bit controllers, you are still forced to access data in at least 32-bits-wide chunks (and that would be 64 bits with DDR-II).

32-bit wide and a burst length of 2, correct? That's 64-bits. Or are there 16-bit DDR chips?

 

[DegustatoR]

I think it's just ridiculous...

What is just ridiculous?

BTW, guys, i'm going to "read only" again. I've got some very... hmmm... surprising new info on NV30, which i won't share w/you, sorry. You'll have to wait for 13 days... (Let's just say, that now i'd add many new question marks to my speclist, posted by alexsok on the first page.)

See you soon.

 

[SteveG]

I think it's just ridiculous...

What is just ridiculous?

read on....

I've got some very... hmmm... surprising new info on NV30, which i won't share w/you, sorry. You'll have to wait for 13 days... (Let's just say, that now i'd add many new question marks to my speclist, posted by alexsok on the first page.)

Umm, you just answered your own question. A "source" that changes his facts from day to day isn't a source at all, just another poser spreading baseless speculation in the guise of informed fact.

 

[Geeforcer]

What is just ridiculous?

BTW, guys, i'm going to "read only" again. I've got some very... hmmm... surprising new info on NV30, which i won't share w/you, sorry. You'll have to wait for 13 days... (Let's just say, that now i'd add many new question marks to my speclist, posted by alexsok on the first page.)

See you soon.

So Nvidia re-designed NV30 in 3 days? Amazing.

 

[noko]

What if Nvidia just post paper specs 13 days from now vice having a real card available to play around with like ATI did for the Radeon 9700 pro? Man, talking about an avalanche of likely arguments over how it will perform. Maybe a good PR method keeping people in mystery over Christmass keeping many from buying from the competition .

 

[martrox]

Bottom line is, so far, we know nothing about this card except rumors and BS!

 

[KimB]

How? With MSAA there is plenty of redundant data to compress (subsamples with same colour).

A lossy algorithm is a possibility, but I'm not sure if I would want that.

The nice thing about a lossy algorithm is that you might be able to guarantee a minimum amount of compression. Then, if you can increase the color depth to compensate for such a lossy algorithm, and still come out ahead, you've won.

The way I see it, this is sort of like the MPEG4 issue. Sure, MPEG4 looks worse than MPEG2, but the cool possibility of MPEG4 down the road is using it to encode higher-resolution videos into less space, resulting in an overall improvement in image quality.