Why do we need tessellation?

[JoeUser]

Im sitting here playing around with this ATI tech demo found here:

http://developer.amd.com/gpu/radeon/pages/RadeonSDKSamplesDocuments.aspx#d3d11 (look for Detail Tessellation 11 (February 2010))

OK...now in playing with this Im noticing that the parallax occlusion mapping looks pretty damn close to tessellation, although with a little more roughness around the edges, but it still has the 3D aspect and depth down pretty damn good. Looks just like actual 3D models. Now I know parallax occlusion mapping can't work well with rounder heads and stuff on character models and objects in a game...but it can certainly work with flatter objects like walls and roads, etc. So why use tessellation with these kind of things in the virtual world? Does parallax occlusion mapping take more processing power?

And then what about tessellation in relation to character models? I mean, why not just use higher polygon count models to begin with? Oh, cause it would be too much of a slow down? Well, thats why models and objects become less detailed the further you get from them (something thats been used forever in 3D gaming).

I don't know...maybe parallax occlusion mapping and higher polygon models use more resource than tessellation...but I don't think thats the case...

So, like the title says...why EXACTLY do we need/use tessellation when there are other methods out there that provide the same results?

 

[evilsofa]

See the HardOCP review of Metro 2033 image quality:

http://www.hardocp.com/article/2010/04/05/metro_2033_directx_11_gameplay_performance_iq/1

Among many other things, it covers what tessellation does for that game, both in terms of image quality and performance cost. Metro 2033 only does it for character models. The summary was that it made character models look very nice with a very low cost in performance.

Looking at an older anandtech article about tessellation back when it was ATI hardware only, the propaganda at that time was that for the same visual effect as higher polygon counts, tessellation was much faster and used much less VRAM.

If tessellation is so beneficial, you might ask why it hasn't been used sooner. Well, they tried to use it sooner. ATI put out its proprietary version of hardware support, and Nvidia came up with its own proprietary version of hardware support (but never actually released it). The two versions were so incompatible that game devs at the time would have been severely inconvenienced to support both, possibly requiring separate executables for ATI and Nvidia, so the game devs collectively said "screw you both, we're not using either until you sort things out", and it took Microsoft putting tessellation into DX11 to break the logjam.

 

[Valset]

to push new hardware what else? but tessellation has plenty of issues, evilsofa list some of them but didn't mention that using it in game causes all kinds of programing headaches. esp with interactions. a good example of this is tessellation mark (otherwise known as Heaven benchmark). look on the stone cobble road with full tessellation on. how would you make something walk on that? this is why you often see its use limited. there are of course ways to do it. I would imagine that this will be less of an issue as time goes on but if you google you will find all sorts of issues with using this in games

 

[evilsofa]

look on the stone cobble road with full tessellation on. how would you make something walk on that?

Ah yes, foot placement. Almost all 3D games use a cheap hack that actually looks pretty good in most cases, and we've all gotten so used to it that we don't notice it anymore. The hack is that placement of feet are not calculated; a simple running animation is used that assumes the model is running on level ground. It looks great on level ground, but not so much on steep hills or steep ramps or especially stairs.

Back when Valve was designing Half-Life 1, they looked at ways of calculating foot placement, but the gaming hardware at the time was not even close to being up to the task, so they put that idea on the back shelf. This is not a trivial problem to solve even now. Edit: apparently I was wrong about it not being used yet - see socK's post below.

 

[kllrnohj]

So, like the title says...why EXACTLY do we need/use tessellation when there are other methods out there that provide the same results?

Because there *AREN'T* other methods that provide the same results. Tessellation generates real polygons, which means lighting works with it just fine, it can be anti-aliased, etc.. It isn't a "hack", so to speak. Parallax mapping is still a flat surface. Flip on wireframe and you'll see it clear as day. The hills are also rougher and the edge is a hard line rather than tessellation's correct edge.

to push new hardware what else? but tessellation has plenty of issues, evilsofa list some of them but didn't mention that using it in game causes all kinds of programing headaches. esp with interactions. a good example of this is tessellation mark (otherwise known as Heaven benchmark). look on the stone cobble road with full tessellation on. how would you make something walk on that? this is why you often see its use limited. there are of course ways to do it. I would imagine that this will be less of an issue as time goes on but if you google you will find all sorts of issues with using this in games

The incredibly bad stonework on the road in Heaven is not the fault of tessellation. That is how they actually designed it. They intended for it to look like that (why I have no idea).

Heaven also completely misuses tessellation. They use it as a complete replacement for polygon detail, when it is more to allow for dynamic LOD and much higher detail. Using it like that means you have no programming headaches from it, no interaction problems, etc...

 

[addictedto60fps]

I don't know about most people, but I really can't tell the difference between tesselation turned on/off in Metro. Maybe it's more pronounced in other titles, but I don't see a difference in Metro.

 

[socK]

Im sitting here playing around with this ATI tech demo found here:

http://developer.amd.com/gpu/radeon/pages/RadeonSDKSamplesDocuments.aspx#d3d11 (look for Detail Tessellation 11 (February 2010))

OK...now in playing with this Im noticing that the parallax occlusion mapping looks pretty damn close to tessellation, although with a little more roughness around the edges, but it still has the 3D aspect and depth down pretty damn good. Looks just like actual 3D models. Now I know parallax occlusion mapping can't work well with rounder heads and stuff on character models and objects in a game...but it can certainly work with flatter objects like walls and roads, etc. So why use tessellation with these kind of things in the virtual world? Does parallax occlusion mapping take more processing power?

And then what about tessellation in relation to character models? I mean, why not just use higher polygon count models to begin with? Oh, cause it would be too much of a slow down? Well, thats why models and objects become less detailed the further you get from them (something thats been used forever in 3D gaming).

I don't know...maybe parallax occlusion mapping and higher polygon models use more resource than tessellation...but I don't think thats the case...

So, like the title says...why EXACTLY do we need/use tessellation when there are other methods out there that provide the same results?

parrallax occlusion mapping isn't without fault, and it could end up being more expensive than the actual geometry, especially on cards with monstrous geometry rates like the newer nvidia ones. i'd guess you could say it can technically be higher quality and is more flexible as well.

Ah yes, foot placement. Almost all 3D games use a cheap hack that actually looks pretty good in most cases, and we've all gotten so used to it that we don't notice it anymore. The hack is that placement of feet are not calculated; a simple running animation is used that assumes the model is running on level ground. It looks great on level ground, but not so much on steep hills or steep ramps or especially stairs.

Back when Valve was designing Half-Life 1, they looked at ways of calculating foot placement, but the gaming hardware at the time was not even close to being up to the task, so they put that idea on the back shelf. This is not a trivial problem to solve even now.

erm, foot inverse kinematics has been in games for years. hl2 does it, doom 3 did it, a ton of games do it now. actually, a game did it the same year as hl1, if i remember right.

 

[Valset]

Because there *AREN'T* other methods that provide the same results. Tessellation generates real polygons, which means lighting works with it just fine, it can be anti-aliased, etc.. It isn't a "hack", so to speak. Parallax mapping is still a flat surface. Flip on wireframe and you'll see it clear as day. The hills are also rougher and the edge is a hard line rather than tessellation's correct edge.



The incredibly bad stonework on the road in Heaven is not the fault of tessellation. That is how they actually designed it. They intended for it to look like that (why I have no idea).

Heaven also completely misuses tessellation. They use it as a complete replacement for polygon detail, when it is more to allow for dynamic LOD and much higher detail. Using it like that means you have no programming headaches from it, no interaction problems, etc...

there are other methods to do this. tessellation is not the final solution by any means. there are dozens of ways to improve the appearance of a game. they may not do it the same way but with the end result being the user experience the fact that they are "real polygons" isn't the point. look at crysis. its approaching 4 years old and is still one of the best looking games. No tessellation. Don't get me wrong I like tessellation but its implementation up to now has left something to be desired

if you go up to evilsofa last post he explains the issue better. and I was using it as a example due to the contrast it provides (heaven that is). the problem remains and has yet to be addressed. interactions with tessellation is an issue. Heaven is just an example.

 

[Mokkat]

I think one of the points of tesselation is that it makes it easier and quicker to build the graphics in a game for the developers. Of course, at the present only a game like Civ 5 is has really done this

 

[Blacklash]

Ya' I've seen the [H]ard Metro article before. I saw nothing in there that would justify a drop from 60FPS to 25 in some cases. DX 11 is often bolted on rather than implemented from the ground up. Most games are developed for the console and a single GPU then ported to the PC. In addition people that do happen to be PC gamers often own a single mainstream GPU. That's where the market is so it's going to have a rather large impact on Dev decisions.

Yes I'm still buying a 6950 or 6970 when they hit

 

[evilsofa]

Ya' I've seen the [H]ard Metro article before. I saw nothing in there that would justify a drop from 60FPS to 25 in some cases.

And I don't see tessellation causing a drop from 60FPS to 25FPS anywhere in that article. Read it again (it's on page 5, "Apples to Apples Comparisons"):

Enabling Tessellation on the Radeon HD 5870 decreased our average framerate by 15%

With the GeForce GTX 480, when we turned Tessellation on, we lost about 7.5% of our average framerate.

 

[Yakk]

Tesselation is ok, and probably if done with it in mind from the ground up, may cut some developer costs and enable better scaling across different performing hardware.

However, for that to work everyone must have access to DX11. Chicken & the egg syndrome. So for now it's a "bolt-on" little bit of eye candy sprinkled here & there on some 3D models.

Other features like Ambient Occlusion Shadows have a much bigger impact right now and cost relatively little to implement.

Y.

 

[ckryan]

The level of tessellation makes a difference in terms of framerates. I don't know what the average tessellation factor is. I know that nVidia's current generation handles higher tessellation factors the ATI, even though radeons have had the ability for 6 generations. I don't know how useful higher factors are now, but liberal use could be greatly beneficial in the future when both developers and hardware have the ability to use it effectively.

Also, the cheapo walking hack is one of the most annoying things I see in games.

 

[Asacolips]

About the walking, that's kind of a moot point. In a game engine, you NEVER use the high detail models for collision detection. Instead, you would use a much lower detail collision mesh that would be significantly easier to calculate for, especially in the light of physics engines and their large number of calculations.

Tessellation's big improvement is that it can increase the model detail with relatively little overhead. It doesn't make a big difference now, because no game has really used it to full effect. If you recall, normal mapping made for a huge jump in detail back when it was first used in games, as it allowed for the game's detail to step up significantly, without a real performance cost. Tessellation would be similar, but primarily for things that need a lot of detail to look right, such as the dragon from the Heaven demo. Until a developer is able to focus on making use of the hi-def models for tessellation, it'll never really reach it's full potential.

Also, it's not a cost cutting measure. Tessellation doesn't magically add detail to the models, it simply divides up a lower resolution model to mimic a higher resolution one: you still have to have an artist make that high resolution model. Well, unless you're just smoothing the model as a whole, but that never looks good.

If developers use tessellation right, which is to make things look extremely detailed up close and to have a good system for LOD as the model recedes away, it would work out awesomely and be well worth the cost of performance. But there's not enough Direct X 11 hardware in people's machines yet, so that may be a while. The year it starts to happen on a large scale, is the year I upgrade to Direct X 11.

 

[SonDa5]

We need it to make UT3 look awesome.


From what I have read the Unreal Engine 4 is one of the main tools that will enable photo realistic games in the next generation of PS4 and Xbox720.

Nice video on how detailed tessellation helps out current unreal engine map making:


http://www.youtube.com/watch?v=IIwnw0VIQDA

 

[Kaldskryke]

I think one of the points of tesselation is that it makes it easier and quicker to build the graphics in a game for the developers. Of course, at the present only a game like Civ 5 is has really done this

Not really. One of the biggest reasons why there are few games using tessellation to any real extent is that making models that tessellate well is not always easy. Games take years to develop and it's not always easy to go redo a whole bunch of art assets. Subdividing existing meshes can often result in rounded corners where they should not be rounded and other problems too. It's not difficult to create tessellatable meshes for terrain, walls, and flat surfaces, but it's a lot trickier for complex meshes. There also seems to be a lack of support for DX11 tessellation in the major modelling software packages which only makes it more difficult for the artists.

OP, POM is an expensive process and it has some downsides. Not only does it suffer from artifacting at oblique angles, but it makes texture filtering a lot more difficult. And if you want a surface to shadow itself with the virtual geometry that's even more computationally expensive. Tessellation has the ability to replace POM with a lot less headaches.

Right now a lot of people have been thinking of Tessellation only as a way to add detail compared to what we're currently used to seeing. It can certainly do a good job of that. I think a much more compelling aspect of Tessellation is its ability to dynamically vary the detail. In general, if an object is close up, tessellate the crap out of it. If the same object is far away, don't tessellate much.

Without tessellation, an artist has to create multiple meshes for each "level of detail" the engine plans on supporting. For example an artist might make 3 meshes named Near, Medium, and Far. I'm not a modeller so I'm going to make up some numbers and say that Near is 50MB, Medium is 5MB, and Far is 500KB.

With tessellation, an artist only creates a single mesh that's about 500KB. The artist also adds a displacement map or other tessellation data which is about 5MB. Not only is this less work for the artist, but it's a memory/bandwidth savings of 50MB, which can really add up. Furthermore the transition between the levels of detail will be smooth instead of discrete. Tessellation even enables the level of detail to vary within the mesh, such as smoothing out the silhouette edges but leaving the middle of the mesh at coarse detail.

Tessellation isn't just about adding detail, it's about making detail more efficient. This unfortunately requires that art assets are created with tessellation in mind from the outset. In the meantime, we have games that try to tack-on tessellation over-top of meshes that are already "detailed enough" in the first place and so the effect is mostly wasted. And considering the none of the consoles support DX11 tessellation it is less likely that any developers are going to adopt this style of art pipeline for a single platform.

 

[Elsie]

Not a very technical answer but HAWX 2 terrain is a great application of tessellation. For me it was nice to look at the edge of the mountain and see smooth curves. For the tiny performance cost I'm totally sold, its not like HDR, which was (for me) difficult to spot, yet had a massive penalty attached. Awful game though, glad I got it for 'free'.