PDA

View Full Version : Without AA/Aniso: NV31 Performance = Ti4600 Performance?


Uttar
12-31-02, 10:28 AM
Hello everyone,

As reported by xbitlabs (http://www.xbitlabs.com/news/story.html?id=1041265473) , the NV31 performance without AA/Aniso is expected to be around Ti4600 levels.

Of course, that doesn't seem too good. But it is, however, a lot better than it seems.

Mufu said in another thread one model will be clocked at 350Mhz and the other one 450Mhz. I've got no idea if that's true, but it seems very logical. So I'll assume it's true.
Thus, the first question is which model got Ti4600 performance. Actually, I think it's slightly more complex than that.
Performance is "comparable" to the Ti4600. And 450Mhz is 128% of 350Mhz
Thus, it would seem logical to say that one model will have performance slightly lower than the Ti4600, and one model will have performance slightly higher than the Ti4600. Let's assume the difference is similar to see what that approximatively means ( it certainly won't be the case )
Then, the 350Mhz model would have 87.5% of the Ti4600 performance and the 450Mhz model would have 112.5% of the Ti4600 performance.

Furthermore, there's Antialiasing and Anisotropic filtering.
I'm assmuing the NV31 will use similar Intellisampling algorithms to the NV30, but with a DDR1 memory system ( DDR2 for a +- $199 part is overkill )
Now, for the aniso hit, we could assume that with 8X, it would be equal than the R300's 16X Performance, with 16X Quality IQ. This seems pretty much right to me. Why? AFAIK, The NV30 algorithm is able to intelligently choose between Trilinear and Bilinear, unlike the R300's. And its algorithm respects IQ more according to nVidia. But then again, all of that might be marketing BS ( let's see that in the NV30 reviews... )

The AA hit, however, is more complex. The NV3x is all about 4X AA. And everything is made with 128 bit in mind. And there's color compression. And so on. So, I guess a hit similar to the Radeon 9700 ( 256 bit bus ) might be reasonable. But it might be slightly less, if Color Compression truly does miracles ( and NV30 reviews will tell that )

So, considering all of that, it sounds like AA/AF performance of the NV31 450Mhz is approximatively similar to the Radeon 9700 performance.
The only difference being it's expected to cost less.

Which is why ATI is releasing the RV350
It sounds like ATI doesn't really want to increase performance with the RV350. It sounds a lot more like Radeon 9700 performance with Radeon 9500 Regular MSRP ( $179 )

What's for nearly for sure now is that the NV31 will be better than every of ATI's current solutions. But the RV350 remains a powerful competitor, and it's hard to say right now if it's going to do its job right.


Uttar

Mod
12-31-02, 11:51 AM
I know that this comment of mine is rather useless, but these speculations are getting too specific... so a bit uninteresting. Well at least for me, sorry. Maybe I should have PM you, I don't know...

No offense from me Uttar, just a comment from someone that still have a TNT2 ultra... :)

Alphy
12-31-02, 12:00 PM
Uttar, I've been meaning to ask you, how CPU dependant do you think these new NV3x cards will be?

Uttar
12-31-02, 12:56 PM
Mod: You don't have to PM me to say that; this is a good way to do it IMO ( of course, some people don't like it, but I'm fine with it )
What type of speculation do you prefer? I agree this is rather boring. Or anyway, it's boring because it's so darn long to say so little. I really should have summarized it or something I guess...

Anyway, any example of the type of thing you find interesting?

Alphy: Well, *very* CPU limited without AA/Aniso. And with 4X AA/8X Aniso, not CPU limited at all :)
Anyway, in future games, Dynamic Branching might potentially make the driver even less of a potential bottleneck ( if used correctly, of course ) - not like it's one with current nVidia cards, but it can be one in CPU-limited games ( such as JK2, IIRC ) for ATI.
That's because it's possible to do the same thing with less DIP calls.
If I understood this correctly, VS 3.0. texture lookup could potentially help this a little too, but I didn't investigate it much.
As I said, this is *very* minor, and taking it into account isn't even very important. The performance gain mostly comes from not having to do state changes.


Uttar

Mod
12-31-02, 04:29 PM
Originally posted by Uttar

What type of speculation do you prefer? I agree this is rather boring. Or anyway, it's boring because it's so darn long to say so little. I really should have summarized it or something I guess...

Uttar

In fact,this is a short speculation, in relation to others. Well I like something more futuristic, because in this way I learn more.

There aren't many variables in consideration in this thread, and so there is almost no room for debate (at least for me).

I would like you to open, for examples, a thread about VS X.X and PS X.X, comenting differences about them, what they are, what we should expect in future versions, etc... :) Well, elaborate some threads about...

Happy New Year :)

Joe DeFuria
12-31-02, 08:50 PM
So, considering all of that, it sounds like AA/AF performance of the NV31 450Mhz is approximatively similar to the Radeon 9700 performance.
The only difference being it's expected to cost less.

I disagree:

Radeon 9700 would have:
1) More fill rate than NV31 at 450/450. (20% more)
2) More bandwidth than NV31 at 450/450 (20% more)
3) Also uses color compression with AA modes...no benefit for NV31 there.
4) Cheaper memory: 275 Mhz vs. 400+ Mhz.

So, I see the price of a 400+ MHz NV31 being about the same as a 275 MHz Radeon 9700. (Radeon 9700 chip would be more expensive, but the NV31 memory would be more expensive.) 9700 is already selling at $230.

I don't see the NV31 performing on par with the 9700. I see it as a notch lower. Probably on par with the 9500 Pro. Better in some cases, worse than others. Also, it's probably more expensive than the 9500 Pro...

For the record, I don't think the NV31 would be 450/450, but max of 400/400. I don't see 450 Mhz memory being cheap enough for the $200 price point. 400 Mhz will probably be pushing it.

jbirney
12-31-02, 10:34 PM
Originally posted by Uttar
What's for nearly for sure now is that the NV31 will be better than every of ATI's current solutions. But the RV350 remains a powerful competitor, and it's hard to say right now if it's going to do its job right.

Normally you have facts and some logic but that above statement really makes me wonder how you hell you came to the above comment....

ASCI Blue
12-31-02, 11:22 PM
I have to agree with jbirney on this one, how will a little gffx compete with the flagship of ATi, that is beyond logic of any kind.

StealthHawk
01-01-03, 02:45 AM
Originally posted by ASCI Blue
I have to agree with jbirney on this one, how will a little gffx compete with the flagship of ATi, that is beyond logic of any kind.

RV350 is not the flagship of ATI. i was under the impression that there would be a R350 and a RV350.

Kruno
01-01-03, 03:11 AM
I'm only intrested in what the cards can do at 320x240 16bpp with everything down and off.

FSAA = not intresting, I barely was able to notice aliasing on my Geforce 3 with QAA on.

AF = Not intrested as the R300 is capable of producing really sharp image. I don't know how much sharper you can get it. My Geforce 3's 4xAF was enough already.
What parts of an image (all 3d image) would bilinear look the same as trilinear?
I have looked at 360 different angles on 1 3d image and all of them looked better with trilinear filtering.

StealthHawk
01-01-03, 06:15 AM
Originally posted by K.I.L.E.R
I'm only intrested in what the cards can do at 320x240 16bpp with everything down and off.

hmm....every card? :p

oh, but some cards require shaders/use nV extensions. so i guess a gf3 is all you need then. please send me your r9700!

MuFu
01-01-03, 10:06 AM
Originally posted by Joe DeFuria
For the record, I don't think the NV31 would be 450/450, but max of 400/400. I don't see 450 Mhz memory being cheap enough for the $200 price point. 400 Mhz will probably be pushing it.

I agree. I think 450MHz is probably the at-speed target for the "Ultra" version of the NV31 ASIC, hence me originally mentioning it. At the time I thought it was using an adaptation of the NV30 memory controller so just presumed the clocks were running synchronously - with 450MHz DDR-II, of course. It doesn't - the NV31 uses a new, DDR-I memory interface. There are quite a few problems with the part right now as well. I'm not sure whether 450MHz is a realistic target anymore.

So yes, what you say is spot on - even 400MHz memory would be cost-prohibitive for a mainstream performance product. They will probably end up clocking the fastest board at 400/350MHz. That would also help to differentiate it slightly from the 5800 - until recently I presumed an "overlap" in clockspeeds between the NV30 and NV31 parts but now it looks like that might not be the case. Depends on how debugging of the crappy NV31 goes, I guess.

MuFu.

StealthHawk
01-01-03, 09:22 PM
Originally posted by MuFu
I agree. I think 450MHz is probably the at-speed target for the "Ultra" version of the NV31 ASIC, hence me originally mentioning it. At the time I thought it was using an adaptation of the NV30 memory controller so just presumed the clocks were running synchronously - with 450MHz DDR-II, of course. It doesn't - the NV31 uses a new, DDR-I memory interface. There are quite a few problems with the part right now as well. I'm not sure whether 450MHz is a realistic target anymore.

So yes, what you say is spot on - even 400MHz memory would be cost-prohibitive for a mainstream performance product. They will probably end up clocking the fastest board at 400/350MHz. That would also help to differentiate it slightly from the 5800 - until recently I presumed an "overlap" in clockspeeds between the NV30 and NV31 parts but now it looks like that might not be the case. Depends on how debugging of the crappy NV31 goes, I guess.

MuFu.

looks like nvidia is having a tough time with hardware all around then.