Final Doom3 benches at HardOCP

Why?

There are architectural reasons that give NV40 a distinct advantage over R420 in D3 style rendering. And, as DemoCoder mentioned, the UltraShadow optimizations are not only suitable for stencil shadow volumes, but for shadow buffers, too.

 

Therer is no doubt that nVidia is going to own DoomIII. As I've said before, DoomIII was designed for NV3x/4x, and the NV3x/4x was built to play DoomIII. There is nothing anyone can do to change this. Bottom line is - no matter what anyone can say - JC favors nVidia, period. That being said, It certianly doesn't help that ATI's OpenGL drivers suck. They (ATI) really do need to do something about it badly. I do believe that the X800 series has much more room for improvement here. Hopefully, there will be massive performace gains with some decient opengl drivers... but theres no way the X800 will overtake the 6800's..... but they should be able to get alot closer, enough to make the differences not "that" important......

I would have thought that ATI could have seen this coming.......

 

Can't disagree with that. If (insert all reasonable conditionals) we'd see a high end TBDR anytime soon ("soon" relative to former conditionals), I'd easily assume that stencil fill-rate is going to be quite a bit higher. I doubt though that anyone could even in that case make any safe predictions on Doom3 performance, based on stencil fill-rates alone; there are quite a few other factors that would also play a role here I figure.

In the case above though I think it's rather a general necessity to have as many Z/stencil units, then just for stencil shadows.

 

I dunno, he may be using a hybrid technique. But it simply doesn't matter, since if he switches to shadow buffers 100%, then zixel performance still matters. There is no scenario I can forsee where having Ultrashadow-like optimizations won't help, that is, in non-AA situations, having double pumping, and the ability to do depth clipping.

That is to say, with the exception of ray-tracing, I don't know of any real-time algorithm for unified lighting that doesn't require multiple passes to calculate either pixels "in the shade" (stencil techniques) or a depth buffer.

 

Benchmarks also say nothing about how the game looks (accurate colors, textures, etc etc).
I'll take a card that gets 20% lower frame rates if whatever game I'm playing looks better than another brand of card. I guess it's all "in the eye of the beholder" but I've liked the way ATI has handled in this regard over the past couple of years.
2D is important 'cause I spent ALOT more time looking at webpages than I do at games.
I'm definately going to give the 6800 a hard look though.

I always played HL in OpenGL. If HL2 is going to be another game to be played in OpenGL, then wouldn't it be safe to assume that these Doom benchmarks could be saying alot about the performance in HL2 as well?

 

Yes, and I've already stated (several times) that we'll see more architectures with increased Z/Stencil. However, I'm not aware of UltraShadow having any uses beyond Stencil Volumes - at least, thats all the GL_EXT_depth_bounds_test extension seems to make note of, as far as I can tell.

 

Well, it seems to be a sledgehammer of a marketing bullet lately...

 

Dave, just out of curiosity, why have you spent so many posts downplaying Nvidia's Ultrashadow / Z/Stencil performance? I think we get your point but as several others have said - those technologies won't be as useless in the future as you claim.

 

<smack> Helloooooo, did you read the thread? Where have I downplayed either UltraShadow or the Z/Stenciling performance for the current generation?? I haven’t – in fact I’m repeatedly stated that I expect other architectures to at least adopt the double pumped Z/Stencil performance in the future. So, how you can construe that as downplaying I don’t know.

What I have said is that I don’t expect future architectures to spend a huge amount more transistors optimizing for stencil performance in the future, beyond what we already have, because – from the discussions I’ve had with numerous people – it’s viewed as a bit of a blind alley ongoing. As LeStoffer points out NVIDIA have been telling developers that shadow maps will be more widely used in the future – this being the case, what would you think it would be more beneficial to optimize for in future?

[edit] Here is the 6800 Leagues presentation on Shadowing techniques. Page 5 echo's LeStoffer's earlier comments, page 22 gives you a big hint about where they are going for future architectures and hence (page 23 and on) what they are optimising for.

 

The NV4x should have zero problems here. Neither should the NV35+. The NV30-34 may produce very slight banding on specular highlights, but we'll have to see.

Well, I can tell you that now that I finally have a good monitor, I can say that the 2D of my GeForce 6800 is every bit as good as the 2D of my previous Radeon 9700 Pro. Right up to 1600x1200 text is crisp and clear. I currently do everything at 1280x960, though, for refresh rate reasons.

The 6800's also have the added benefit of digital vibrance, which can help games somewhat, but really helps movies, I think.

I was pretty sure that HL2 was going to be a purely Direct3D game.

 

1) Having to enable Glslang support through registry hacks, non WHQL drivers through leaks, then having compatability problems, one such thread..many more and much more on Nvnews.

http://www.beyond3d.com/forum/viewtopic.php?t=11358&start=0


2) 3Dlabs picks up rendering errors: http://www.beyond3d.com/forum/viewtopic.php?t=11712&start=0

The [H] article shows a 9800XT outperforming a 5950, even with Ultrashadow ??

http://www2.hardocp.com/article.html?art=NjQyLDQ=

 

The depth bounds test is a general method you can use whenever you do an additional pass that might affect the surfaces already laid out in the frame/depth-buffer. If you do it right, you can use this for shadow buffers as well.

 

It certainly should be enabled by default. I don't understand what's the issue with "non WHQL drivers through leaks".
This thread shows one issue with an older driver, and one which doesn't seem to be related to GLSL at all. There are more recent drivers, you know?

So what? The compiler compiles every valid GLSL shader just fine.
And it extends GLSL.

Without AA. And R360 has other architectural advantages over NV38.
NV40 vs. R420 is a completely different story.

Hmm hmm.

 

Humus posted his Demos on Nvnews, I have two kids that want to head to the beach or I would post the links, but the results were not pretty.

The Demos that Humus wrote almost always had rendering errors on a FX on those threads. I get a sense that now the Nv40 is here, the FX doesn't exist anymore

Hmmm Hmmm ??

 

I'd like to emphasise what ailuros is saying about the stencilling capabilities - in relation to current chips. A Kyro 2 performed on par with the much faster Geforce 4 (clock 250+ vs 175mhz), larger (~60mil vs ~15mil trans & 4x2 vs 2x1), much greater bandwidth (250ddr+ vs 175sdr) and 2-generations further advanced - with stencils - as tested by Fablemark.

Modern high-end chips have over 10x as many transistors & are clocked much faster (some about 3x faster). A high-end tbdr carrying through such capabilities would surely surpass even a double-pumped imr architecture in handling the stencils of doom engine graphics. Imagine about 30 times the z-capabilities of the Kyro in a modern high-end tbdr architecture (ie about 30x faster 'z' than geforce 4)- in handling the stencils - im sure mr carmack would enjoy such a speed-up in doom 3??? (maybe its a slightly exaggerated estimate, with probably more emphasis on shaders, but who knows???)

The Geforce 6800 ultra with its double pumped z-units would struggle to offer even half that kind of theoretical z-performance for stencilling- but what do others think???

Obviously the complete architecture would come into play- but starting out with over a 2x stencilling advantage from the best current doom 3 card - would be a fair start.

 

Please define "D3-style" rendering...

 

Odd, though, that probably 90% or better of 3d gaming is D3d... Interesting disparity, certainly. I would also imagine that 90%+ of the sales of 6800 and x800-class cards are made and will continue to be made to people whose primary use of them is 3d gaming.

Because of the lop-sided numerical disparity between D3d game engines and OpenGL game engines, it's always been much easier for me to understand ATi's inclination to concentrate on D3d. What's puzzled me for a long time is nVidia's seeming inability to compete on the D3d driver side of the equation, considering the relative ubiquity of D3d game engines of myriad description. Does this mean nVidia is a "one game-engine pony"? (Nah, just kidding, sort of...)

 

Stencil shadows with "Carmack's Reverse" algorithm and attenuated lights, which means:
- an initial Z pass
- a stencil pass per light
- flipping the Z test mid-frame
- shadow volume capping
- light sources only affecting a certain depth range
- fast stencil test will help a lot

All these things NV40 can do faster than R420.

 

Thought I would let you guys know what is coming this next week as I know some of you were not happy with how narrow the first results were.

We did a LOT of tests with a lot of CPUs and Video Cards. And we DID use the GOLD MASTER of the game for testing.

GF3, TI4600, 5800, 5900, 6800, 680GT, 6800U are all coming this week. As well as 8500, 9600XT, 9800Pro, X800Pro, X800XT. There are probably a couple I missed in there as well.

Here are the systems and CPUs we used the cards on. (Not all cards on all systems, but there is some overlap.)

 

sweet