NV35 - already working, over twice as fast as NV30?

[noko]

Lets blow the smoke out of the way.

Quake III Max settings
1600x1200x32
ATI drivers 4xAA, 8xAF Performance (who knows what Nvidia used for AF settings, probably the crap aggressive settings.)

R300 at 250/250(500) core/mem(ddr)
88 FPS

R300 at 250/320(640)
98.2FPS

R300 at 250/350(700)
100FPS

My XP was running at 1.76ghz for these tests as a side note. Looks like core limited situation with memory above 320mhz.

Now wasn't the NV35 getting 110fps at 1600x1200x32? unless I misread the leaked info? Also my next questions is, isn't ATI's 4xAA much better then Nvidia 4xAA? Meaning to me that the NV35 isn't no killer, in addition the tests is really showing the speed increase due to memory bandwidth increase due to 256bit memory bus and really doesn't show anything about the real speed increase due to other core enhancements.

I beginning to think too much has been assumed and extracted from these leaked results. Sounds like a test to show the advantages of increase memory bandwidth in best case situation only for the NV35. Nothing more I think can be accurately taken from the data.

 

[Me]

And then there's the R390.
Uttar

R390? First I've heard of it. I thought the next jump for ATI was R400. It hardly seems to make sense doing another re-spin...

 

[Arun]

And then there's the R390.
Uttar

R390? First I've heard of it. I thought the next jump for ATI was R400. It hardly seems to make sense doing another re-spin...

According to several generally reliable people, it would seem that ATI delayed the R400 until next year, and the R390 ( on 0.13, probably ) is the product they'll use to try keep the performance crown until next year.

The reason for the R400 delay, according to those same people, could be that the current process technology wouldn't allow such a product at a reasonable cost.

There are rumors of a MAXX R390 too ( MAXX = ATI's name for 2 chips on one board solutions )
*plug* Read www.notforidiots.com/GPURW.php to stay updated on those type of things */plug*


Uttar

 

[Mephisto]

The reason for the R400 delay, according to those same people, could be that the current process technology wouldn't allow such a product at a reasonable cost.

I highly doubt that. The situation in the manufacturing area is well known since quite a while. Sounds more like feature creepness or design flaws rather than unmature process technology.

 

[Arun]

I highly doubt that. The situation in the manufacturing area is well known since quite a while. Sounds more like feature creepness or design flaws rather than unmature process technology.

Possibly. Those people aren't exactly sure on that point, but are sure that it's delayed.

But remember that the R400 was targetted to use Low K. And it sounds like TSMC is having more and more trouble with it, and it won't be ready for quite a while. Could be a reason for ATI to delay the R400.
And that wasn't know for *so* long - there were serious rumors that suggested nVidia was still planning on Low K for the NV35 in late Q4 2002!


Uttar

 

[Evildeus]

@noko Don't forget that the chip was running @ 250MHz and not 500 MHz

 

[Tahir2]

@noko Don't forget that the chip was running @ 250MHz and not 500 MHz

He didnt forget that, that's why he benched at 250MHz Core and then 250MHz DDR Memory plus his variations which shows that the QuakeIII results are more smoke and mirrors from NVIDIA.

 

[Pete]

ED, my last post made it pretty clear that the card was running at 250/1000 (to use the vernacular).

noko, thanks for your post. Interesting that nV isn't much faster on a game they traditionally have the edge on. Now we get to wonder which AF setting nV was using.

 

[Tahir2]

Now I'm confused too... :?

 

[Evildeus]

@noko Don't forget that the chip was running @ 250MHz and not 500 MHz

He didnt forget that, that's why he benched at 250MHz Core and then 250MHz DDR Memory plus his variations which shows that the QuakeIII results are more smoke and mirrors from NVIDIA.

Well, i think he does, and even more, he says:

Looks like core limited situation with memory above 320mhz.

How much does he gain when the chip is clocked @ 320MHz?

That would be interesting to see how much the NV35 could gain.

 

[Evildeus]

What the hell, downclock to 250mhz so that you can compare speeds :?, sounds like Nvidia is having problems with higher NV35 core speeds otherwise wouldn't they have a slightly faster core comparison to begin with? Or is that the speed where the difference between the NV30 and NV35 is maxed out .

I don't know why, but Mufu already stated that the chip was running @ 400MHz so...

 

[Tahir2]

Yes, but now we need to or should compare the numbers to the NV30 results to make the comparison valid since we know the core is the limiting factor at >320MHz memory.

noko also discusses this that when you pair the memory at full speed with a core at 250MHz we get similar results..... (sorry figuring this out in my head)... so the NV35 per clock cycle on the core is finally either able to compete to R300 per clock cycle or exceed it.

So er.. the core has been improved/drivers are better/some bugs were eliminated/NV35 has true 8 pipelines unlike NV30.

... Can someone explain what we can extrapolate from the results because I have a MAJOR HEADACHE (for real).

 

[Evildeus]

Yes, but the consummer will pay the final product and for him it doesn't matter if x is faster clock per clock than y, if at final y is faster than x.

Like the R300 is already faster clock per clock than the NV30!

 

[Tahir2]

I am the consumer as well..
The consumer will care if the product is good or bad and vote with his/her wallet.

But I am not talking about that. I am talking about what we are meant to make of noko's results and what you mean by "Don't forget the chip was running @ 250MHz and not 500MHz"

I'm asking because maybe I am missing something here..

 

[Evildeus]

Well, i think his post is contruary to what he wants to prove.

Look the R300 have the same figures but are limited by the core, so the NV35 @ 250 MHz must not be forgotten!

 

[Arun]

Actually, there's another possible explanation to this, Tahir.

There's obviously some memory interface in both NV35 and R300.

The R300 generally got memory *slower* than the core.
The NV30 got memory at the same speed as the core. We don't know for sure with the NV35, but we think that it'll either be as fast or faster.

If the memory interface isn't able to keep the memory busy, then it might sound like it's memory limited even though it's core limited.

And as shown by the above facts, the NV35 got a more efficient per core clock memory interface.
Thus, the NV35's advantage could very well come from here!


That is, if I'm not too tired and that I'm not making completely false assumptions and the above post isn't completely lame and incorrect.


Uttar

 

[Pete]

What are you two (ED & Tahir) smoking?

noko's numbers show that the R300 at 250/700 is almost as fast as an NV35 at 250/1000 at about the same settings (assuming the NV35 is running Balanced AF).

All the NV35 vs. NV30 comparison shows to me, ATM, is that NV35 does indeed have a 256-bit memory bus and some slight core/crossbar improvements (to explain the improvement that's greater than the mere memory bandwidth bump).

It's meant to be a core-limited situation, to emphasize NV35's fancy new memory bus.

MuFu stated in NVNews that NV35 may be meant to be run at 400MHz DDR, but the test we have before us is at 500MHz--again, to emphasize memory performance.

I'm not sure how noko's post is "contrary to what he wants to prove." His results show that the R300 is achieving near-NV35 speeds at 250MHz core. They also show that the core is the limiting factor at 16x12x32 w/AA+AF above a certain memory speed (~300MHz, in R300's case), thus NV35 also has "eight true pipes" (or, in nV's terms, the ability to output eight color values per clock, double NV30) to make use of its extra memory bandwidth.

In short, we're all crazy, and nV doesn't cheat on 3DM03.

 

[Evildeus]

Well if it wasn"t core limited, he would have a point. But as he pointed himself, the result on the R300 are core limited and that just show that it's the same as for the NV35. The 250MHz is still important and more than we think with noko results as a basis of comparaison!

 

[Tahir2]

noko's numbers show that the R300 at 250/700 is almost as fast as an NV35 at 250/1000 at about the same settings (assuming the NV35 is running Balanced AF).

That's what I thought but at settings we don't know of, of course apart from 1600*1200*32 4*AA..we don't know the AF level for example, tri or bilinear etc etc ad nauseam.

The only reason that NVIDIA released these or got leaked is that NV35 will be faster than the NV30 but they dont want to give up too many surprises (like that NV35 was meant to be faster.. d'oh!) otherwise NVIDIA jeopardize the GFFX Ultra even more so. I was just getting more and more confused as to what was going on after every post.

What are you two (ED & Tahir) smoking?

Golden Virginia. Smoking is bad for your health. Want some?

 

[MuFu]

But 250/500MHz with DDR-I is insane. I very much doubt they've been able to achieve that.

MuFu.