NVIDIA's detailed slam of 3DMark03

[Chalnoth]

Originally posted by kyleb
but that assumes that no one ever finds a beter way to do things than the method you are cacluateing. that just comes off like a silly think to assume for me.
There are obviously many factors, but the idea is very simple. If the circuit size approaches the wavelength of light of the frequency the curcuit is operating at, it radiates.

Granted, processors are highly-complex devices, and it may be possible to reduce the amount of EM radiation by intelligent circuit routing, but that can only go so far.

The main point here is that very high frequencies are going to be much, much more challenging to attain than more parallelism. Thus, if companies focus on more parallelism, they'll be able to put out overall faster processors.

[StealthHawk]

Originally posted by Chalnoth
If I remember correctly, the Kyro II got such high scores because it didn't render the cubemaps that the other cards had to deal with.

hmm, you could be right about that. Kyro2 doesn't support cube maps AFAIK. Sweeney complained about that a long while back.

[StealthHawk]

Originally posted by kyleb
well game engines are not very good without video cards are there? so could you please explain to me how you seperate the two?

3dmark03 vs 3dmark2001. in 3dmark03 people with the same video card but with different CPU speeds have almost the same score. not true of 3dmark2001.

first off, i dont not see any evednece to back your claim that the gf2Ultra was way ahead of the Kyro2 when it came out. as i recal i considerd them both as fairly equal options when my voodoo5 was falling behind and i made choice to get a gefoce3.

http://www6.tomshardware.com/graphic/20010425/kyroII-12.html

kyro2 loses almost all tests to the original gf2. kyro2 wins some at high resolution and 32bit. the gf2Ultra was a good 30% than the gf2 in such situations. so extrapolate the results yourself.

anyway, look at more than one page of benchmarks from the link you posted; the only thing amazing i am seeing is how amazingly loose your definition of "fact" is. how does that one page prove the point you were atempting to make any less than benchmarks on other pages disprove the same point?

that kyro2 wins at all is the point. although as Chalnoth has suggested, this could be due to cubemaps not being used.

can i get a clear answer on this questions before going into the rest of it?

you want it, you got it.

[jbirney]

The question of weather or not the games will be more/less CPU dependent is a true unknown. If the game is a multi-player only then the likely hood that its more CPU limited is smaller then it beeing held back by other bottlenecks. Any single player game has a higher chance of being CPU limited. Lots of them can be CPU limited if you tweak the settings (like take UT2k3 and turn up the phyics detial to high and you stress your CPU more).

Also dont forgot some of these timedemos that they use to record benchmarks (JK2 for example) are used with bots = more of a CPU limitation.

UT2k3 is less CPU limited than UT ever was. Remember UT was basically the Unreal Engine with bot support and a few tweaks. It was design for software rendering. And that software rendering stole much of the CPUs power. So that is ONE game that has moved away from being CPU limited. Which makes me happy as know I can work on getting my AI stuff more complicated to make the bots "fight" better :)

Personally I think we will see a mix. Some games will use more of your CPU others less :)

[kyleb]

exactly jbirney, and i recal Sweeny bitching about cube maps on the kryo as well so i stand by Chalnoth argument there.

[StealthHawk]

Originally posted by jbirney
The question of weather or not the games will be more/less CPU dependent is a true unknown. If the game is a multi-player only then the likely hood that its more CPU limited is smaller then it beeing held back by other bottlenecks. Any single player game has a higher chance of being CPU limited. Lots of them can be CPU limited if you tweak the settings (like take UT2k3 and turn up the phyics detial to high and you stress your CPU more).

hmm, but the botmatch scores in UT2003 ARE more CPU limited than the other timedemos, doesn't that mean in multiplayer UT2003 becomes more CPU limited?

Also dont forgot some of these timedemos that they use to record benchmarks (JK2 for example) are used with bots = more of a CPU limitation.

that explains a lot.

UT2k3 is less CPU limited than UT ever was. Remember UT was basically the Unreal Engine with bot support and a few tweaks. It was design for software rendering. And that software rendering stole much of the CPUs power. So that is ONE game that has moved away from being CPU limited. Which makes me happy as know I can work on getting my AI stuff more complicated to make the bots "fight" better :)

true enough. but the original UT was a bad example of a 3d accelerated game. the engine was coded the way it was because Sweeney did not forsee the advent of 3d accelerators.

Personally I think we will see a mix. Some games will use more of your CPU others less :)

undoubtedly true. the question is how much of a mix. will we see many games that go to the extent of 3dmark03? will that be the norm? or the exception to the rule?

these are the questions that need to be answered as 3dmark03 is "the gamer's benchmark."

[kyleb]

StealthHawk, i sweer you must have taken lessons from johnny cochran in the oj trial.

[Myrmecophagavir]

Originally posted by StealthHawk
hmm, but the botmatch scores in UT2003 ARE more CPU limited than the other timedemos, doesn't that mean in multiplayer UT2003 becomes more CPU limited?The bots need lots of CPU time for AI, that's completely removed in multiplayer.

true enough. but the original UT was a bad example of a 3d accelerated game. the engine was coded the way it was because Sweeney did not forsee the advent of 3d accelerators.UT's Glide support was very good. I'd say he just didn't foresee hardware T&L.

[StealthHawk]

Originally posted by Myrmecophagavir
The bots need lots of CPU time for AI, that's completely removed in multiplayer.

is the botmatch benchmark a recorded demo or is AI actually being calculated?

this question came up in an Intel vs AMD thread and nothing conclusive ever came out of it.

actually, everyone but me said it was just a recorded demo with no AI in it at all.

so....which is correct?

[kyleb]

yes, ai is acutaly calculated in the benchmark, and they are all on the unreal setting asweel. so unless you play a bunch of unreal bots all the time, the benchmark is prety syntetic as well. although it is still a nice little benchmark in its own right.

[Chalnoth]

Originally posted by StealthHawk
hmm, but the botmatch scores in UT2003 ARE more CPU limited than the other timedemos, doesn't that mean in multiplayer UT2003 becomes more CPU limited?
Not necessarily. The question remains as to where those additional CPU cycles come from. Do they come from just drawing the bots? Or is there a whole lot of AI being calculated?

After all, if a very significant part is just the AI, then that part will never be used in multiplayer (since it will all be done server-side). In fact, regardless of where the limitations lie, the game will be somewhat less CPU-limited in multiplayer than it would be in a botmatch.

[Chalnoth]

Originally posted by Myrmecophagavir
UT's Glide support was very good. I'd say he just didn't foresee hardware T&L.
Nope. If you pay attention to how the game was rendered in 3D, it totally does not make sense for any 3D accelerator. The benefit with Glide was basically that Sweeney was able to code more "to the metal," and thus avoid many of the stalls that the engine caused for other accelerators.

Said another way, Glide just allowed Sweeney to avoid the inefficiencies he put it in the focus for software rendering. It's not that the engine supported Glide well, it's that Glide was well-suited for the engine.

[jbirney]

Originally posted by StealthHawk
is the botmatch benchmark a recorded demo or is AI actually being calculated?

this question came up in an Intel vs AMD thread and nothing conclusive ever came out of it.


Hmm I have the source code at home so I can check tonight for ya...

[kyleb]

Originally posted by kyleb
yes, ai is acutaly calculated in the benchmark, and they are all on the unreal setting as well. so unless you play a bunch of unreal bots all the time, the benchmark is prety syntetic as well. although it is still a nice little benchmark in its own right.



!!!!

[Chalnoth]

Yes, this means that the botmatch scores are also much more unreliable than the flyby scores. In other words, whenever using botmatch scores, make sure you take a few runs. And whenever looking at them, expect that the benchmarker did not, and so the scores can vary by a few fps and still be within the statistical error of the benchmark.

The flyby scores appear to be highly repeatable by contrast, and make an excellent video card benchmark (whereas the botmatch benches make for more of a CPU benchmark).