NV3X subpixel precision in 3DMark03

[MuFu]

Nah, I think it'd look more like this:

Interesting now how the initial NV35 excitement has died down now and things have descended into lunacy. We seem to be waiting on further clarification regarding these sparkling pixel issues and also the apparent FSAA post-filtering.

If the former were due to a drop down to integer-based vertex ops then surely it would not manifest itself in GT1, a DX7 test?

MuFu.

 

[demalion]

No, that's what Wavey will look like with the increased caffeine intake and non stop testing in the near future.

Not that such a concern should stop us from pestering him anyways, of course.

 

[John Reynolds]

Pfft, you kidding? He already looks like that.

8)

 

[digitalwanderer]

Damn, what's going on with thread layout these days?!

If it's not giant blue teapots it's huge white blocks all over the place.

"Sometimes you look into the darkness, and sometimes the darkness looks back."

Me likes the two mirrors for a demo idea, infinite regression always holds an endless fascination for me.

Sooo...has anyone tried renaming 3dm2k3 to something else and running the benchy on an FX?

 

[MuFu]

Sooo...has anyone tried renaming 3dm2k3 to something else and running the benchy on an FX?

What, like "Tridentareabunchofcheaters.exe"? (click for article)

MuFu.

 

[MuFu]

Maybe the nV hacks involve pre-emptive occlusion.

The path of the camera in the benchmarks is fixed, so it'd be quite easy to "hard-code" extremely efficient hidden-surface removal for each benchmark.

What's worrying is that if they can do such a thing for 3dm, they can do it for timedemos too.

MuFu.

 

[BoardBonobo]

That would be where you'd get your motion antialiasing from then, you know the spatial position of each element over time. The two main ingredients for maa, spatial and temporal.

 

[MuFu]

BINGO.

http://www.extremetech.com/article2/0,3973,1086025,00.asp

Damn, I was just in the middle of writing to MikeC of nV News asking whether it'd be ok to post a thread with my "thoughts" on the matter!

MuFu.

 

[Joe DeFuria]

Wow...MuFu gets a cookie!

 

[MuFu]

No cookie... too late.

Stupid, party-pooping ExtremeTech.

MuFu.

 

[Joe DeFuria]

You know, if this is all true (that nVidia is purposely taking advantage of pre-determined view frustum to reduce overdraw...)

Can you imagine the Doom3 benchmarks, using an nVidia created time-demo?

On this note, I wonder how long nVidia had with the iD created timedemo before they actually allowed benchmarks....

 

[MuFu]

Yeah - we'll have to wait and see. UT2003 Flyby/Botmatch is the other "big one".

Must be a few people shitting themselves at nV right now.

MuFu.

 

[RussSchultz]

On this note, I wonder how long nVidia had with the iD created timedemo before they actually allowed benchmarks....

Not enough time to do what you're suggesting. Or are you also suggesting that they sneak in every time John tests things to mislead him into saying his famous "NV30 path is faster than R300 in ARB2"?

 

[antlers]

On this note, I wonder how long nVidia had with the iD created timedemo before they actually allowed benchmarks....

Not enough time to do what you're suggesting. Or are you also suggesting that they sneak in every time John tests things to mislead him into saying his famous "NV30 path is faster than R300 in ARB2"?

I'm not sure they would need very much time at all to create a "cheat" driver if that was what they were intent on doing. They've had access to basic information about Doom 3's engine for some time. For all we know they already had an instrumented driver that captures the information they need to make a custom driver with "optimizations" for a certain application/camera path. Depending on how automated they've made the process, they could probably churn out a driver customized to support a specific timedemo in a few hours.

Once a company has shown a determination to cheat in elaborate ways, it's difficult to put anything past them. They certainly have a lot of development resources to apply to the problem.

I'm not saying that the whole observed advantage of the NV3X architecture with the NV30 path in Doom 3 over the R3xx is artificial. But it is certainly conceivable that the driver/benchmark made available by NVidia exaggerated the difference.

 

[jb]

Keep in mind they had enough time to build a custom map and a timedemo

 

[RussSchultz]

Keep in mind they had enough time to build a custom map and a timedemo

Right, but its tough to implement the sort of alleged cheat when the map and timedemo change on you. Its also tough to fool the developer who's a pretty smart cookie and works with his engine daily (presumably running different things than the set timedemo).

 

[Joe DeFuria]

Not enough time to do what you're suggesting.

See antlers4's post.

Or are you also suggesting that they sneak in every time John tests things to mislead him into saying his famous "NV30 path is faster than R300 in ARB2"?

Odd, I thought his famous quote was actually: "At the moment, the NV30 is slightly faster on most scenes in Doom than the R300, but I can still find some scenes where the R300 pulls a little bit ahead. "

In fact, I'm sure of it.

In any case, I was also under the impression from Carmack that the ARB2 path did indeed actually WORK on NV30, which for "some reason" Tom Pabst was unable to verify...

 

[jb]

Right, but its tough to implement the sort of alleged cheat when the map and timedemo change on you. Its also tough to fool the developer who's a pretty smart cookie and works with his engine daily (presumably running different things than the set timedemo).

Russ,

they were showing Doom3 scores on their custom map at Comdex during the NV30 launch back in Nov/Dec 2002. Thats plenty of time.

 

[RussSchultz]

Odd, I thought his famous quote was actually: "At the moment, the NV30 is slightly faster on most scenes in Doom than the R300, but I can still find some scenes where the R300 pulls a little bit ahead. "

In fact, I'm sure of it.

You win. You got the quote right. How does that change anything, though? Even ignoring the recent HardOCP/Anandtech results, it still appears that the cards are in parity when it comes to Doom3 _coming from the developers mouth_ who we can assume has hardware and runs everything through enough hoops to find this particular cheat.

In any case, I was also under the impression from Carmack that the ARB2 path did indeed actually WORK on NV30, which for "some reason" Tom Pabst was unable to verify

So now Carmack is involved in the conspiracy? (Though, I think Mr. Pabst was doing his benchmarking on an alpha release of Doom3--not sure though and not sure how exactly that'd affect a cards ability to run a particular path or not)

 

[RussSchultz]

Right, but its tough to implement the sort of alleged cheat when the map and timedemo change on you. Its also tough to fool the developer who's a pretty smart cookie and works with his engine daily (presumably running different things than the set timedemo).

Russ,

they were showing Doom3 scores on their custom map at Comdex during the NV30 launch back in Nov/Dec 2002. Thats plenty of time.

And this map was the one used for the results on HardOCP and Anandtech? No.