-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Unigine DirectX 11 benchmark - Tesselation porn
- Thread starter -SD-
- Start date
brain_stew
Member
Inkwell said:
Tesselation is geometry though. That's why its so much better than normal maps and the like, no trickery just real, proper geometary is generated, so it looks great from every angle, doesn't have unsightly artefacts and interacts properly with lighting and shadowing systems.
That's a far too simplistic view of rendering. I can't prove it of course, but I'm willing to bet that much (the majority) of this hit is not actually caused by the tesselation operation itself but rather by the amount of very small triangles on screen causing a setup rate limitation and/or massively reducing the efficiency of the actual shading (because the individual tris are much smaller than 4x4 blocks).M3d10n said:
brain_stew
Member
M3d10n said:
Perhaps, it could also indicate that the dedicated hardware is a bottleneck and that the GPU's general purpose hardware could do a better job of it. See: ATI removing their texture interpolaters from RV870 because they were bottlenecking performance.
Edit: And durante explained another possible scenario, fact is, we just won't know until Nvidia release their hardware and there is much more widespread uses of tessellation. Even then we might not know exactly what is going on.
Look closely at the textures in this shot:Inkwell said:
http://unigine.com/press-releases/091001-directx11/house_no_tesselation.jpg
Pay attention to the brick texture and you'll notice it does have the details for the sides on it, since it is drawn more or less like this:
Code:
_____
|\___/|
| |_| |
|/___\|That's another option. Or any combination of theseM3d10n said:
. I really hope someone does an in-depth performance analysis of this once more DX11 GPUs are out.This is a good point, I shouldn't have called it a height map in my earlier posts, it really has 3 components.M3d10n said:
No - it just changes the storage for model-bumps from vertex to texture(in simplest form). Someone still has to make the bumps.Danthrax said:
And increases aliasing artifacts about ten-fold. It's good the new GPUs can throw so much resources at AA nowadays...brain_stew said:
Minsc
Gold Member
Durante said:Bububut it's a hardware tesselator! There can't be any significant slowdown!
(Sorry, just poking fun a bit at some people in the Fermi thread
Heh, fair enough. I'm shocked to see over a 50% drop in framerate, but naive I suppose, this certainly is not the first feature to share such a fate. Even AvP previews were talking about the tessellation is working without creating any hitch in the playability, which I figured they were speaking from actually running the game. Maybe ATI's tessellation talk is all smoke and mirrors after all... it may very well end up completely impractical in games without Crossfire, but like you said, until we see how it all stacks up, it's hard to judge.
Software could result in a 90% drop in framerate for the same results for all we know. It's disappointing either way to see that this technology takes such a toll on performance, and will probably end up another graveyard feature of gaming.
brain_stew
Member
Fafalada said:
Well I'd rather have aliasing that can be dealt with by standard msaa hardware than the increasing amount of "texture/shader/specular" aliasing present in games with heavy use of normal maps are increasingly showing. At least I can deal with that aliasing without resorting to bruteforce supersampling.
Although ultimately, we're going to have to make that break away from reliance upon standard msaa someday it seems. There's just so many more things in a modern game engine that can cause aliasing artefacts these days that throwing an ever increasing amount of multisamples at the problem really isn't going to help you all that much in the end.
What are your thought's on Epic's touted "analytical AA" is their a future in something like that? I've started to see some pretty decent implementations of shader based "edge AA" recently (even if it isn't a true alternative to standard msaa), and best of all the performance impact never seems to be too great either.
Oh, ok. That makes sense.wmat said:
I was under the impression that the engine generates the geometry. Unless I'm wrong here as well, I was trying to say that I'm surprised how good it looked for being generated by the engine and not hand sculpted. I still may be wrong. Also, I worded that original post horribly. Sorry about that.brain_stew said:
That makes sense, but it still seems like the texture may be stretched a little during tesselation, or compressed when not tesselated. It just seems like with the way the brick is drawn, the "side" of it is much larger when it is tesselated.M3d10n said:
MrJollyLivesNextDoor
Member
brain_stew said:
think i've got a decent understanding of this, i've used displacement mapping in 3dsmax - a height map is applied to a polygonal surface to produce geometry similar to how a normal/bump map simulates geometry.
One thing i don't understand, if this is a simple case of adding additional poly's using maps and tesselation, whats stopping these characters/objects being modelled with similar detail/poly counts in the first place? Are there other benefits i'm not seeing?
Also how does this affect LOD, rather than having 3 or 4 preset models swapped in depending on when the detail is needed i guess tesselation could reduce poly counts on the fly?
Jamesearlcash
Member
Wow, makes things stick out a bit more.
Oh, don't get me wrong: I actually don't expect a hit like that in real games. This is a benchmark designed exclusively to show off tesselation/displacement mapping so it uses it (IMHO) even for stuff that a normal game would likely use real geometry for. I also think the uptake for it will be better than any other recently introduced DX feature -- though that's admittedly not saying much.Minsc said:
I agree with this. On the other hand, if you ahve tons of poly edges the advantage of MSAA compared to SSAA will be reduced (to 0 in the degenerate casebrain_stew said:
).Well, you already touched on one thing, it's much easier to do dynamic LOD (eg. just fade out the displacement at some distance). But the most important thing is that you can do the displacement on the GPU, so you only need to send the low(er) detail models to it (and also only animate those). The vertex data for raw models at this detail level can get huge.MrJollyLivesNextDoor said:
The Hermit
Member
Damn, why I didn´t buy a DirectX 11 card
brain_stew
Member
Pimpbaa said:
No.
Can someone with a 58xx take a screenshot with the dragon statue in a similar position to the first one by DennisK4? I thought the tesselation for the flagstones around that was particularly well done -- not over the top like other parts of the video but still greatly improving the overall impression.
brain_stew
Member
MrJollyLivesNextDoor said:
Yes, this is actually one of the major benefits of dynamic tesselation, no more jarring LOD pops (in theory anyway, and dependant upon implementation), an ever increasing amount of geometry is added to models as the camera gets closer in an analog fashion.
The basic advantage of tessellation? Performance, basically. Sure the benchmark seemingly has as much as a 50% reduction in framerate with tessellation enabled but it also shows several orders of magnitude increase in geometric detail as well. Try creating a scene on modern hardware with traditional forward rendered rasterisation with no tessellation routine and that many triangles and performance won't drop by 50%, it'll probably struggle to manage a single frame a second! When the camera is zoomed in, in that demo, you're seeing models with triangles that are sub pixel in size, even at high resolutions common on PCs these days. That's just a huge quantum leap over the poly count of today's games.
sotodefonk
Member
I like how the tessalation stuff, cast shadows too... anyone who has tested it with a dx11 card... how is the performance hit?
MrJollyLivesNextDoor
Member
brain_stew said:
thanks, explains a lot.
I was a little confused with the reasoning behind this being to simplify thing for performance yet performance loss was the result - seems it's not an apples to apples comparison.
I've thought polygon detail is one of the few things that have been ignored in recent years in relation to shader complexity and texturing, plenty of real geometry keeps the whole scene looking much more detailed when used right imo.
red shoe paul
Banned
Milk Lizard said:
Agreed. I was expecting my mind to be blown just a tad, guess not. Specially with that house. It's almost laughable. :lol
Assuming the typical use will be adaptive tesselation, I agree. If not - it won't be easily solvable by MSAA either.brain_stew said:
Like Durante mentions, high polygon density makes these methods along with MSAA basically obsolete (well, MSAA just becomes SSAA). But until we get to the point where all polygons are pixel sized, they definately have their use.
It's probably safe to say that all next-gen consoles will support some form of GPU tesselation, yes. (Well, who knows what Nintendo is planning)Sushen said:
Yeah, I wonder what will happen in terms of AA in that future (with many polys <= 1 pixel). Of course, that's not the only thing that would have to change drastically. With the way rendering is done currently you'd also basically have most of your shading hardware idling in that case. Of course, both of these can be "solved" by just throwing ridiculous levels of SSAA at them -- which also allows GPU companies to sell their stuff well into the future :lolFafalada said:
BuddhaRockstar
Member
That stone road would be hell on your feet.
Mr. Durden
Banned
I love the tech, but it really fucks up the texture work. I hope it's just because they want to show the "big" difference, and that it doesn't represend what it'll do when used it actual games.
Similar threads
- Locked