A promise is a promise :) there you go!

Just throwing out a wacky idea, but couldn't it be possible that the 9800Pro/XT could have more than one thing as it's limiting factor? And that this would (partially) explain why we're having trouble comparing the test vegetto talks about with the one x-bit has done?

There are so many unknown variables even when looking at the game alone, that saying the NV40 numbers are impossible since a 9800Pro/XT example can't explain the behaviour, is a bit strange to me. (Not to mention that that 9800pro and 9800XT has small differences between them which might in part explain some of it).

So why not try to understand what the limiting factors of the game is, before starting the fan-war that is currently building momentum?

Here is what the UT2k4 release notes says about the topic:

Also, since we don't know what settings the nv40 was run at, this might be relevant:

Anyways, my point is, even though I like a good nvidia bashing every now and again, that it might be more worthwhile first to figure out what has changed from 2k3 to 2k4 and then use that to figure out how accurate the numbers we've gotten are, instead of just attacking the guy/numbers based on the previous version of the game.

 

About the UT2004 botmatch results:

Was the proper method for botmatch benchmarking followed? Some discussion.
Is it even necessary, or maybe other things are necessary?...UT 2004 isn't necessarily exactly the same as UT 2003 in such matters.
What map was used, and where are the comparisons for it using currently released cards? <-This seems to be the heart of the running argument about the botmatch results.

I have a lot of questions related to my understanding of the UnrealEngine's adaptation of workloads. As one example, what if more detailed models are used at higher resolutions? A higher LOD determination and polygon count for the engine's purposes at a given distance, with fixed function vertex processing, would seem to result in a host system workload for animation and skinning of models.

However, all that really seems to concern is the last in the above list of questions and why it is necessary for it to be answered...it seems that it should be clear that a 1600x1200 4xAF and 4xAA botmatch results could very easily be GPU limited on such a fast host system, and that these results are possible.

...

It seems confirmed that it is indeed 16 pipelines in fixed function (a good thing), sort of delivering what was hoped for from the Volari V8 Duo. The remaining questions are concerning shader performance and efficiency in functionality.

If nVidia has had anything to do with recommendations for the various benchmarks seen or intended the body of programs the card doesn't work with, this would seem to strongly indicate that floating point processing parallelism is not the same. Without looking to some exotic design properties, the transistor budget would seem to indicate the same thing. Considering nVidia's history of communication (which would seem more than slightly likely to accompany a sample card) and driver behaviors, I think this is the case.

However, since I expect a minimum parallelism of 8 for floating point processing, my only concern here for the NV40 being a (very) good high end card (moving forward) is that it has that minimum for PS 3.0. What I do wonder is if the PS 1.3 parallelism might be 16 (which goes with some teasings from a little while ago). I think this is very likely as well...integer dependency for performance seems to be a comittment with nVidia's devrel discussion "treat as DX 8". Also, 16 TMUs seems established with fixed function, and "undoubled up" (or "unstalled") in relation to the NV3x designs seems to allow it. With a parallelism of 8 for floating point processing, this makes a lot more sense than the NV30, even for sharing similar philosophy, as long as issues such as floating point render targets are addressed and register issues aren't made worse.

This would also indicate the shading power quotes (that seem provided by nVidia?) might be significantly more accurate than what nVidia has provided before, depending on the shader functionality.

As for AA and AF, it seems perverse that we actually have direct screenshots of this and we're still left with so many questions. The available aniso and AA testers are not secrets by any means, so was the person proscribed from using them by NDA, recommendation, courtesy or such? Or just didn't think of it?

 

Not to be rude or anything, but isn't saying that "since I couldn't overclock my cpu, he can't overclock his cpu" a bit... far fetched?

Had you on the other hand known what exact model and batch the testers cpu came from and you had another from the same batch, you might stand on a better ground, but even there it would still imo be .... far fetched.

Also, we know nothing of what he has done to achieve this overclock. What cooling system is he using, did he overlclock by connecting L1 bridges, etc.

So unless you can show me a larger empirical base of evidence showing that the A64 3200+ can not be run at 11, it shouldn't be taken for granted that what you are experiencing is the norm.

 

I'm not clear if the UT2K4 Demo numbers at Xbit are flyby or botmatch, but I believe you can only get botmatch numbers from the demo. As such, Xbit got 70fps out of the 9800XT @ 10x7 AA+AF and 95fps @ 10x7 plain, so the CPU does allow for numbers as high as Vegetto claims.

In short, those UT2K4 demo Xbit numbers don't seem to disprove Vegetto's claims. What am I missing?

I see the 128-bit 9800SE (3120MP/s, 11.2GB/s) is slower than the 128-bit 9600XT (2500MP/s, 11.8GB/s) with AA+AF, indicating to me that perhaps memory bandwidth is more important than anything else (fillrate, vertex shader power) at those settings. The 9600XT is 7% faster at 10x7, a slightly greater disparity than its 5% greater memory bandwidth, but it drops to just 5% faster at 12x10, practically equal to its memory bandwidth advantage. I'm probably overlooking other factors, so this line of reasoning is relegated to a potentially flawed afterthought.

And I don't mean to appear to pick on you, jvd, it's just that your quotes were the closest at hand. I'm just having trouble understanding the disbelief of the UT2K4 numbers over the other ones.

 

according to www.overclockers.com the _average_ overclocking result from this cpu is 2429 mhz. So yes you are correct that this chip can be overclocked MrG.

later,
epic

 

I agree with you their pete but because of JVD's strong pressure on this the poster changed the hardware to a dual configuration which makes me cast alot of doubts. I believe some of the details could be true but they might not be as close to first hand as originally stated and other details made up. Then again it could be a totally lie.

I belive JVD's problems with the number where problem's with XBIT's numbers rather then Vegetto's.

 

If it's totally CPU limited,how come we can get much higher FPS in 640x480 than in 1600x1200? I really have problem to understand it.

The very easy way to determine how a factor affects the final result is to reduce other factors' effect as best as we can,provided there's no correlation between them. And I think what FPS you get at 640x480 (not at 1600x1200) is closer to how fast the CPU can calculate AI/Physics,etc.

 

I totally agree, when we benchmark CPU's we do it at low resolution as to not let the GPU hinder the score.

So if I am right in my thinking the more powerfull the card the higher the 1600x1200 score would go.

 

You NOOB's outa do yar homework he was NOT referencing, or debating that the 3200+ CAN Overclock to 2420 Only that your all dumb enough to think it will do 11x multiplier in the process

There is NO KNOWN method to enable multipliers 11+ the 3200+ A64 is only able to use multipliers 10 and less.... Do your homework or shut yar mouths

 

Well I'm no nvidia fan boy.
I'd even say I am an ATI fan boy.

I just don't buy the 'evidence' that ut2k4 is CPU bound in a way that means no next gen gpu (be it ATI or nVidia) can get better than 37fps on that system.
Especially when a fair bit of the 'evidence' is ut2k3.
& especially when there are questions about the similarity or lack thereof between the systems.

Look at the Max Payne 2 tests for a really cpu bound test.
In that test, its only at the extreme end where it just starts to get gpu limited & that the oced 9800pro falls behind the xt.
So infact, if anything the 37.2/37 arguement is proving that ut2k4 is actually gpu bound for that test.

If ut2k4 is cpu bound it should be showing it at the low res, no aa/af end not the top end.
<looks>
Ooh, actually it does show up there as all the cards have generally higher scores than later on.
But even there its still mostly gpu limited as there is a clear variance between cards.
I'd say (based on fudge work & comparison with max payne2 results) the fps limit would be something like 100fps which makes 87fps at 1600 with aa/af quite plausible given a next generation top end card.

 

Well thats is a reasonable abusive first post I hope your 2nd post isn't like that.

Just for the record you can change the FSB and I've seen reports over 2500 Mhz which is well above 2420.

 

Wow, you sure read really well dont you. Did i ever post that it was possible to enable 11+ multiplier?? No, what I said was that its possible to get to 2420+. I should have clarified that I think Veggeto might have gotten confused as to how the overclocking was done, since it not his computer. He has gotten the specs of the machine wrong before, so I thought that he might have gotten this part wrong too. Btw

later,
epic

 

haha now justify, and make up an imaginary cover story

 

Oh, and the A64 doesn't have a Front Side Bus, just for the record

its called HTT, and yes that CAN be changed, multiplier is locked though, and I doubt it was mistaken specs More likely not quite enough knowledge to get the facts straight, but enough to fool most of you

 

At least your second post was more civil. Again I ask you if I ever said/wrote/posted that it was possible to get a multiplier of 11+. I was only trying to say that the overclock was possible.

later,
epic

 

Quite true, I was the one who talked about the 11x multiplier...

But NewbAnnihilator, since you seem to know this, has anyone mapped out all the different multipliers available by connecting different L1 bridges? (At least I seem to recall its the L1 bridges which controls the multiplier)

I'm not talking about setting a higher clock multiplier in the BIOS, but rather modifying the cpu so that it thinks it's using an 11x multiplier.

Anyways, welcome to the forum.

 

Well just to make it C L E A R it was MrGaribaldi who thinks the 3200+ will run at 11

 

You seem incapable to answer a simple question in a matter that you know so much about....

But to put it in easier terms:

Has anyone, that you've heard of, tried to connect the L1 bridges in different ways to see if they could change the multiplier?

 

Welp Nice BIG hypothetical IF To bad there are NO visible bridges that have been, or could be cut That leaves only the possibility of a pin trick, and even then its far more likely that AMD has done as Intel, and locked the multiplier on die. So considering all that, I find it highly unlikely, and most likely, impossible, this fellow with the fictidious Nvidia has figured out a way to do this. Believing any of this without even a single shred of evidence is just hilarious.

 

I just did you silly person, I think you outa go back and watch tripping the rift, suite you knowledge level much more.... Bridges on the A64