Post NV30 related news here!

[jbirney]

If I remember well, NVidia stated implicetely nv30 would not use 256 bits, announced the tape out (end of design); ATI launched the R300, we heard of unexpected delays for nv30 and the last thing we know... bing! It is there! Looks like the architecture team has been busy during those unexpected delays...

Very very very unlikely. There are soo many things that would have to change if you jumped from 128 to 256 memory bus (like packaging, pin out, PCB routing, the crossbar controllers would have to change meaning more or grow from a 32 bit to a 64 bit ect).

nV30 is done. About the only thing they can do know is tweak the fab process to get the speed, yeilds or minor bugs fixed.

[StealthHawk]

here's what i found most interesting in that articleNV30 is different from its predecessors because it focuses more on the GPU rather than the memory. The shift means that the performance bottleneck shifts from memory efficiency to computing efficiency. In addition to this the NV30 includes better algorithms for detecting when hidden pixels that applications are not rendered through the system. This is known as an extension to Hidden Surface Removal and will be a part of a new system NVIDIA said we couldn't talk about, at this time.

we've been bandwidth limited for such a long time, ever since the gf1

[Lezmaka]

To me, it looks like ATI is doing something similar to what nvidia did with the 4200. Only the 4600 & 4400 were annonced/released, then a month or two later, we saw the 4200 show up.

Maybe this time, we'll get the nv30 equivalent of all 3 at once.

[Remi]

Originally posted by jbirney
Very very very unlikely. There are soo many things that would have to change if you jumped from 128 to 256 memory bus (like packaging, pin out, PCB routing, the crossbar controllers would have to change meaning more or grow from a 32 bit to a 64 bit ect). You're probably right, it is just that.... as their architecture group seems to study a lot of things in advance and as they have several teams working simultaneously on different projects, I would not be that surprized if the right solutions were already almost fully designed when they decided to put them into silicium... In addition, those guys certainly do know how to do their job fast.
Just speculations however... not based on any other facts than that string of events. If someone have a better explanation, sharing is certainly welcome! :)

[DadGT]

Can someone please post the link to the David Kirk interview where he talks about 256 bit interfaces being "overkill"? I am interested in the exact wording he used.

[Waltz King]

:) Good luck, it's not in English:
Kirk Interview (http://www.hwupgrade.it/articoli/603/2.html)

[SavagePaladin]

I'm actually hoping they only release two versions of the NV30...
budget, which is a non crippled yet non supremely fast version, and regular
As it is I'd rather get a 4400 because it isn't crippled, and doesn't cost my legs.

[DadGT]

Thanks for the link.

It's pretty much what I thought. Currently, a number of people don't believe NV30 can be 256 bit memory because David Kirk said he thought it was excessive. People interpretted this as "NV30 will be 128 bit." But from google and babblefish translations, David Kirk actually said he thought it was excessive "at the moment." But what is "a moment" to nVidia? Also, remember that this interview was right after Parhelia was announced with 256 bit memory interface. We see where that chips performs compared to the 128 bit GF4 line.

If NV30 is 256 bit memory (and it's still rumors to me until NV actually announces), I see absolutly no contradictions from Kirk from this interview.

[Remi]

Good point, "at the moment" might mean on .15 process for instance... What's your opinion?

[StealthHawk]

Originally posted by Remi
Good point, "at the moment" might mean on .15 process for instance... What's your opinion?

"at the moment" could also have been used to talk about the pseudo DX9 Parhelia. aka, it might not be farfetched for a true DX9 part. obviously these "cinematic effects in real time" will gobble a lot of bandwidth

[DadGT]

I know David Kirk isn't usually giving many hints about real future products. I think his "at the moment" statement really was about what was out or really close to being out. I don't think he said NV30 will not be 256 bits at all.

Slightly related to this. I went back to look at the Parhelia specs. Am I missing something or is the 256 bit interfaceof the Parhelia not muliplexed? By that I mean something like nvidia's crossbar where the 128 bit is actually 4x32 bits. I know ATI went multiplex with R300. Seems this could play a significant role in effective memory bandwidth use.

[StealthHawk]

Originally posted by DadGT
I know David Kirk isn't usually giving many hints about real future products. I think his "at the moment" statement really was about what was out or really close to being out. I don't think he said NV30 will not be 256 bits at all.

Slightly related to this. I went back to look at the Parhelia specs. Am I missing something or is the 256 bit interfaceof the Parhelia not muliplexed? By that I mean something like nvidia's crossbar where the 128 bit is actually 4x32 bits. I know ATI went multiplex with R300. Seems this could play a significant role in effective memory bandwidth use.

yup, Matrox seems to have used the brute force approach with Parhelia. i also suspect that their drivers aren't as optimized as they could be.

[jbirney]

Correct,

there are few optimization in the Parhelia MB. However keep in mind that it has as lower pixel fill rate which does not really help it too much. The fact it beats a GF4 with HighRes, Max FSAA and AF shows that it does have some power their. But not the most efficent use of bandwidth at all.


You can take David's words in a number of different ways. I also will see if I can get a english version of that review as there were plenty of them floating around. Also keep in mind that 256 bit memory != 256 bit address bus. Again from reading a few different reviews where it was really easy to see that he did not tink where a memory like that was needed.