NV30: What we *do* know

[Uttar]

Hello everyone,

The recent R300 vs NV30 article showed us one thing: we do a know a lot about the NV30 if we look very deep into nVidia documents.

And thus, i propose to begin a thread which shall be centered on certain information. Things which are highly unlikely to be incorrect.

So, let's begin by what the R300 vs NV30 article proofs

1. The NV30 will have 3 Vertex Shader units
2. More instructions for VS/PS than the R300 in every front, but less temporary registers in the VS ( only 16, the R300 got 32 )
3. No direct MRT support ( But possible using pack/unpack. However, that's harder to use for programmers! )
4. 100-120M transistors ( recent investor documents saying 100+ gives doubts to everyone )
5. AGP 8X ( Interesting, eh? )

I'm *not* convinced by the author's 8 pipeline justification ( however, i agree i fail to see how they'd get to 120M with 6 pipelines... ) - and since this is for certain information, none of that here! :)

http://developer.nvidia.com/docs/IO/3260/ATT/OpenGLforNV30.pdf

Page 8, NV_vertex_program2 Performance
First of all, let me say the author of the R300 vs N30 article didn't undertsand the "Older VP1.0 and VP1.1 programs go faster too" statement

All that means, IMHO, is that the three Vertex Shader units will also be used for VP1.0 and VP1.1 - they're only confirming there won't be driver bugs...

However, another statement is a lot more interesting:
"Less overhead launching vertices"

There are a lot of explanations for that. They could simply have optimized overheads or they could be compressing them.

No matter what's the used system, i'd say that truly justifies the "over" in "Over 1.5x" and "Over 3x"


Okay, so anyone found something else?


Uttar

[Kruno]

Depending on the scenario probably a 25-50% increase in performance over the R300.

Lets talk bandwidth and fillrate first :)

[Lezmaka]

Originally posted by K.I.L.E.R
Depending on the scenario probably a 25-50% increase in performance over the R300.

Lets talk bandwidth and fillrate first :)

We won't know anything about bandwidth or fillrate until we know clock speeds and if it's 128 or 256 bit bus. And those we probably won't know until it's launched, or very close to it.

[Uttar]

Bandwidth: No tile-based rendering ( see David Kirk interviews ).
70% chance that it's a 128-bit bus.
Color & Z Compression ( like the R300 )

Fillrate:
8 pipelines
More instructions, resulting in possible lower instruction number for same operations compared to the R300.
Higher possible instruction number, enabling very high quality for specific spots in the image.


Uttar

[DadGT]

Reading over the recent Beyond3D r300/nv30 comparison article made me realize one thing, we don't "know" jack about the nv30 chip. We do know a lot about the functions it implements and the some microarchitectural details of registers. But we know very little of the chip itself.

I wonder, since the 80486 and P4 support the same x86 instructions (minus SIMD), does that mean they should perform similiar? Making performance assesments based on what instructions a chip supports seems very speculative to me.

Until we know what the product nv30 will actually be, I don't think any real conclusions can be made.

[Uttar]

Agreed that we can't make any performance guess with the little info we got on the chip itself.

However, that doesn't mean we can't try to find other info.


Uttar

[Fotis]

We only know that nv30 will have 120million trans ,use DDRII on 0.13 and wont be called geforce.

[Uttar]

Originally posted by Fotis
We only know that nv30 will have 120million trans ,use DDRII on 0.13 and wont be called geforce.

Actually, i disagree with 2 of your 4 points.

DDRII: Yeah, we know that from the Inquirer. Now it gotta be right! :rolleyes:

Won't be called a GeForce: We are not certain of that. What we DO know is that it won't be called "GeForce 5"

Recent rumors suggest GeForce FX. But then again, those are nothing but rumors.


Uttar

[DadGT]

I will question the 120 million transitors also. Didn't that 120 million number come from an interview? IIRC, in that same interview, it was stated NV28 would have 70 million transistors. That wasn't true. Why should the 120 million number be correct.

We DO know (from nVidia's financial presentation which I do believe) that NV30 will have 100+ million transitors. There is one fact. Of course 500 million is greater than 100 million, so the financial statement reveals no real details.

[Uttar]

Originally posted by DadGT
I will question the 120 million transitors also. Didn't that 120 million number come from an interview? IIRC, in that same interview, it was stated NV28 would have 70 million transistors. That wasn't true. Why should the 120 million number be correct.

We DO know (from nVidia's financial presentation which I do believe) that NV30 will have 100+ million transitors. There is one fact. Of course 500 million is greater than 100 million, so the financial statement reveals no real details.

Actually, in my original post, i said between 100-120M, just didn't feel like complaining more :)

However, i'd still bet it will be 120M

That number comes from Jen Hsun Huang. He said in a Wired Magazine interview. He did not speak about NV28/NV18 there.

However, he also said 120M in an investor conference where he said 80M for the NV28. But i guess plans change. Pretty much the same thing happened with the GF3 Ti 500 - many sources told it was going to be a modified core but, in the end, it wasn't because nVidia changed their plans.


Uttar

[StealthHawk]

Originally posted by Uttar
However, he also said 120M in an investor conference where he said 80M for the NV28. But i guess plans change. Pretty much the same thing happened with the GF3 Ti 500 - many sources told it was going to be a modified core but, in the end, it wasn't because nVidia changed their plans.


Uttar

which is exactly why the transistor count, pipeline number, and other "givens" like no geforce 5 name may mean nothing.

[DadGT]

It's been said before, throw your net wide enough and you are bound to catch a few fish. It's basically impossible to separate what is rumor from what is knowledgable rumor. I stopped trusting anything as concrete a long time ago.

[volt]

Brian Burke NVIDIA's Desktop PR manager on a question about ATI's GDDR-3:

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.

My money is on DDR2 @ 500 ;)

[volt]

Originally posted by Uttar
...DDRII: Yeah, we know that from the Inquirer. Now it gotta be right! :rolleyes:

Uttar

Not only that but it has been indirectly stated by Brian from NVIDIA.

[Uttar]

StealthHawk, i agree with you that nothing is 100% certain. That's why i said "which are highly unlikely to be incorrect" in my original post :)

And now that you're stating that, i'm really doubting the number of Vertex Units ( supposed to be 3 )

For that you see, in that document which says "Over 1.5x faster than NV25" , you also got a disclaimer which says:
"NV_vertex_program2 & NV_fragment_program likely to be subject to minor changes"

So, for developers currently developping on a platform ( NV30 emulation ) which is really, really slow... what would be the smallest change?
Performance of the final system.

So, i'm really wondering if that disclaimer doesn't simply mean that performance estimations they give are incorrect. But then again they don't give performance indications for NV_fragment_program. Maybe there'll be other minor changes too, or maybe i'm 100% wrong.


Anyway, now that i see that quote again, i disagree with you volt.

My money's on GDDR2 @ 500, not simple DDR2 :)
But then again, he could simply be trying to make us hope.

However, here's a part of the quote you didn't say:

"Problem is, what ATI calls GDDR-3 is quite similar to what Nvidia, Samsung and others..."
Hmm, isn't the rumor that it's Samsung DDR2? Hmm...

Oh wait, this was a fact thread? Damnit :P Now, since i'm the poster of the orignal thread, well, i suppose i've got to punish myself

"Tsk tsk! This is a FACT thread! Bad myself!"

...


Uttar

[imtim83]

So i would think no one knows the 100% for sure offical NV30 specs till Nvidia offically releases them. I have been taking everything as rumors because you can just never know.

[StealthHawk]

Originally posted by Uttar
My money's on GDDR2 @ 500, not simple DDR2 :)
But then again, he could simply be trying to make us hope.

what's the difference :confused: does GDDR2 have less latency or something?

i really doubt that the NV30 will have 3 VS. that's an odd number, pun intended.

besides, all competitors that even remotely have some DX9 compatibility all have 4. i don't see why nvidia would fall behind in this.

[Uttar]

I've got no idea of the differences between DDR2 and GDDR2

However, my best guess is that they found some type of optimization for memory bandwidth but it couldn't work with DDR2. At all.

That could also be one of the explanation for delays ( beside 0.13 TSMC problems ) - they hoped it would work on DDR2, but for a reason or another, it simply didn't. So they had to create their own memory specification and put it in production, which took quite some time.

But then again, this is only speculation...


Uttar

[Juntari]

Since NV30 is speculated to have 128-bit memory bus, does that mean that its overall memory bandwidth is similar to what R9700Pro could muster? :confused:

And since the PR guy quipped that 500MHz is boring in 2003, would that mean that faster memory is being planned to phase into NV30?

From the aforementioned article, it seems that NV30 has more in line with the latest DX9 revision features than with R300. Could those be possibly useful and implemented until the next cycle?

OR quite simply 25% to 50% improvement over R300 could be the most reasonable speculation.

[StealthHawk]

Originally posted by Juntari
Since NV30 is speculated to have 128-bit memory bus, does that mean that its overall memory bandwidth is similar to what R9700Pro could muster? :confused:

no. if it has a 128bit bus it will have less raw bandiwidth than the R9700Pro, which has ~20GB. NV30 with 128bit bus and 500MHz ddrII would have ~16GB. of course this discounts optimizations like z compression, color compression, the fabled HSR NV30 supposedly has, etc.

And since the PR guy quipped that 500MHz is boring in 2003, would that mean that faster memory is being planned to phase into NV30?

probably in a refresh of NV30.

[legion88]

Originally posted by Juntari
Since NV30 is speculated to have 128-bit memory bus, does that mean that its overall memory bandwidth is similar to what R9700Pro could muster? :confused:



No. You need two variables to calculate the bandwidth: one is the memory bus width and the second is the memory frequency.

Bus width is 128-bit but you have no information on actual memory frequency.

[imtim83]

No one knows for 110% sure what the NV30 specs will be correct? Also no one knows for 110% sure if the NV30 will be 128 bit, 256 bit, etc correct? I am taking everything as rumors about the NV30 so far. I remember a rumor about the ATI 9700 pro being one month late but it was not. I did not beleive it either and it wasn't true at all either. Just saying its you should not believe anything for 100% sure till nvidia releases the offical NV30 specs and when Nvidia says it will offically be out. I don't favor either. I like Nvidia and ATI both equally.

[Kruno]

HTF can a probability be over 1?

Originally posted by imtim83
No one knows for 110% sure what the NV30 specs will be correct? Also no one knows for 110% sure if the NV30 will be 128 bit, 256 bit, etc correct? I am taking everything as rumors about the NV30 so far. I remember a rumor about the ATI 9700 pro being one month late but it was not. I did not beleive it either and it wasn't true at all either. Just saying its you should not believe anything for 100% sure till nvidia releases the offical NV30 specs and when Nvidia says it will offically be out. I don't favor either. I like Nvidia and ATI both equally.

[Philibob]

Originally posted by K.I.L.E.R
HTF can a probability be over 1?
:D
It was right once ;)
Just saying its you should not believe anything for 100% sure till nvidia releases the offical NV30 specs

[ben6]

that response I got from BB did seem to indicate 500MHz DDR-2 this year didn't it :) as to the memory bus , I ain't commenting, though I assume , if NVIDIA went 8x2 pixel/texture units they would need all the bandwidth they can get.