Crysis 2 Tessellation Article

[3DVagabond]

I think the key is why did one of the worlds most impressive engineering institutions feel that higher levels of tessellation are needed?

It just may -- because Crysis 2 uses more tessellation than any title to date and yet when one reduces the tessellation, there are artifacts.

Maybe if the developers used x16, it would of been full of artifacts and not worth adding at all. I would like to know the answer to this.

I'd have to see it before I could honestly try and answer that, and I'm not likely to. A very likely cause though is texture mapping errors. Since textures are applied to a model per polygon. It's easier to mess up UV coordinates when editing a model than it is to keep it right. The assumption that it's because of 16x tessellation though makes no sense at all. At least none I can think of. You can subdivide a model as much as you want to without ill effects. Each time you subdivide it, it gets smoother, nothing else, at least once you've weighted the vertices. In this case though we have the game applying 64x and then after the fact, software trying to reduce it. It would be very easy to mess up the model during this process as opposed to not applying as much in the first place.

Of course, the equally ridiculous appeal to authority is why did one of the world's most impressive engineering institutions feel that tessellation is only effective up to 16x?

AMD has stated that anything more than one (tri)poly per 6 pixels is ineffective. Sorry, it's from an old article. I doubt I could find it. Hopefully you can accept my word that it's true. Or, can remember it as well.

 

[SirPauly]

I'd have to see it before I could honestly try and answer that, and I'm not likely to. A very likely cause though is texture mapping errors. Since textures are applied to a model per polygon. It's easier to mess up UV coordinates when editing a model than it is to keep it right. The assumption that it's because of 16x tessellation though makes no sense at all. At least none I can think of. You can subdivide a model as much as you want to without ill effects. Each time you subdivide it, it gets smoother, nothing else, at least once you've weighted the vertices. In this case though we have the game applying 64x and then after the fact, software trying to reduce it. It would be very easy to mess up the model during this process as opposed to not applying as much in the first place.

I don't know. Your theory could be plausible to me.

 

[SirPauly]

Here's a link with a breakdown of the performance hits. I've posted this link before.
http://www.behardware.com/articles/...ser-look-at-performance-and-tessellation.html

There's also a graph that shows what the Ultra performance hit is with AMD's tessellation slider set to 0.

The part that is odd is the percentage hit is less with a 5770; why?

More logically, the Radeon HD 6800s benefit from having a larger tessellation unit buffer to gain in efficiency over the Radeon HD 5800s. The Radeon HD 5770 actually has an identical tessellation unit to the Radeon HD 5870, but benefits here from having fewer processing units to reduce the bottleneck in terms of access to the memory shared by these units.

Interesting.

 

[3DVagabond]

The part that is odd is the percentage hit is less with a 5770; why?.

It varies from 57% to 52%, depending on card. It's 1% different than the 6800's. The 5870 and 6900's take the biggest hit. The most likely culprit is driver optimizations.

We all know how the performance hit happens. AMD uses tessellation hardware that has a finite amount of resources available. Once it's exceeded performance goes in the tank. nVidia's tessellation can just call upon more resources. AMD's POV is that there's no IQ reason to subdivide a model beyond 1tri/6pixels. My experience with game modeling would agree with that. Actually, that's more resolution than is really needed.

AMD could likely make a gpu with the same level of tessellation power. It would just as likely be as big and as inefficient as Fermi is though. On 28nm they might be able to get that much computational power in a reasonably sized part though. Who knows, but it seems like they might. While I agree we don't need it for gaming, it should be useful in other apps.

You know, all of this compute power isn't free. It's costing the consumers money. Look at who's benefiting from it? It's not you or me that is getting any benefit from needing ~$500+ video cards to truly enjoy a game. It allows the game devs to be lazy and the card manufacturers to sell us higher profit hardware.

 

[Silverforce11]

Or you could not buy Crysis 2 and waste your time with a crap console port. Crytek does not deserve my money.

 

[SirPauly]

Actually the title is pretty good though and the graphics are impressive. I realize there is a bigger hit with AMD but if one still can enjoy the tessellation and simply drop shaders and post processing down a notch. I have to tweak my sabotaged GTX 470 as well, and another strength if one has S3d, the hit is so small to use it.

It was a good offering but didn't raise the bar like the original did, but there are goodies.

 

[Keysplayr]

Or you could not buy Crysis 2 and waste your time with a crap console port. Crytek does not deserve my money.

It's actually a pretty good game. And quite nice looking to boot.

 

[Firestorm007]

It's actually a pretty good game. And quite nice looking to boot.

Agreed. It's a good game and the price has come down to make it a no brainer. I play it on my 6970 just fine.

 

[NoQuarter]

You may sing a different tune when AMD has more powerful tessellation and more is added in future titles, one may imagine.

No one should be losing performance because of poor implementation of a feature though.

This youtube video shows a lot more objects and it looks like a lot of them they just broke the large polygons into thousands of tiny ones without actually adding any more information to the model. On many of them you can still see the way a ton of triangles just form 1 large flat triangle from the old model. There are clearly some spots where you can see the tessellation was used to deform and add depth though.

Also this John Carmack interview about tessellation was pretty interesting, and he points out tessellation is only really useful if you built from the ground up with it in mind.

 

[Seero]

No one should be losing performance because of poor implementation of a feature though.

This youtube video shows a lot more objects and it looks like a lot of them they just broke the large polygons into thousands of tiny ones without actually adding any more information to the model. On many of them you can still see the way a ton of triangles just form 1 large flat triangle from the old model. There are clearly some spots where you can see the tessellation was used to deform and add depth though.

Also this John Carmack interview about tessellation was pretty interesting, and he points out tessellation is only really useful if you built from the ground up with it in mind.

"Tessellation is only good with micro polygon size" said John Carmack, from the link, which is true. What is not true is what AMD claims that polygon should not be smaller than 16 pixels with tessellation.

Suppose John is someone who knows a thing or two about graphics , than there is nothing wrong with the first link as polygons are pretty micro IMO. Following what John said, he stated that current generation of video card do not have the power to utilize tessellation and generate reasonable FPS, which appears to be nailing this thread if you use Crysis 2 as a sample. The optimization is simply tessellate some objects, but not all of them. In other words, Crysis 2 is being conservative on tessellation already as current gen video card won't be able to run it in full details.

 

[3DVagabond]

"Tessellation is only good with micro polygon size" said John Carmack, from the link, which is true. What is not true is what AMD claims that polygon should not be smaller than 16 pixels with tessellation.

Suppose John is someone who knows a thing or two about graphics , than there is nothing wrong with the first link as polygons are pretty micro IMO. Following what John said, he stated that current generation of video card do not have the power to utilize tessellation and generate reasonable FPS, which appears to be nailing this thread if you use Crysis 2 as a sample. The optimization is simply tessellate some objects, but not all of them. In other words, Crysis 2 is being conservative on tessellation already as current gen video card won't be able to run it in full details.

It was 6 pixels, not 16 and do you have the link? Or, can you point me to the post that has it. I would like to read his reasoning, thanks. Never mind.

 

[mosox]

The optimization is simply tessellate some objects, but not all of them. In other words, Crysis 2 is being conservative on tessellation already as current gen video card won't be able to run it in full details.

it's not like that. You can 16-pixel tessellate every single surface in a game and the current video cards will cope with this. Subpixel tessellate only a few surfaces (and leave all the rest non tessellated) and the cards will struggle because a large amount of normal tessellation will not create a bottleneck while a small amount of hypertessellation will.

Crysis 2 is tesselated exactly the way you do it when you want to create bottlenecks and cripple the video cards.

The overtessellated objects are not even important and a gamer won't even look at those barriers or sidewalk borders let alone the water you can't see.

 

[3DVagabond]

double post

 

[3DVagabond]

"Tessellation is only good with micro polygon size"

That quote is so pulled out of context. As well as not being an accurate quote.

He's comparing the level of detail that you can get from normal maps, which is one pixel. He's defending the lack of tessellation in their game. He states that tessellation is at it's best at sub pixel resolution, not that it is only good there. He says that because normal maps can't go finer than 1 pixel and tessellation can. They've chosen to use normal maps because the models really should be designed from the start for tessellation and you can't get better than normal map detail from it anyway because we don't currently have the GPU power to go there.

He says other stuff to defend their lack of tessellation, but it would be better for folks to listen to the interview themselves. It'll save misquoting and added spin.

 

[houe]

Tessellation is a buzz word and a scam.

 

[3DVagabond]

Tessellation is a buzz word and a scam.

I think it will be a great tool. Once games/models are designed from the start to use it. As of right now? You are mostly correct. I say mostly because, used intelligently, it could add detail to the game in ways that other techniques find it more difficult. Until consoles use Dx11 though, it's not likely to happen.

A bit off topic here, if I may. For some reason people on this forum have a real love affair with game devs and producers. Most of them are scummy though. They'd rip your old grandmother off for her last dime. Trust me, they'd develop something for a game that served no noble or artistic purpose. All it requires is they get paid for it.

 

[SirPauly]

it's not like that. You can 16-pixel tessellate every single surface in a game and the current video cards will cope with this. Subpixel tessellate only a few surfaces (and leave all the rest non tessellated) and the cards will struggle because a large amount of normal tessellation will not create a bottleneck while a small amount of hypertessellation will.

Crysis 2 is tesselated exactly the way you do it when you want to create bottlenecks and cripple the video cards.

The overtessellated objects are not even important and a gamer won't even look at those barriers or sidewalk borders let alone the water you can't see.

Thankfully, there are many more areas that offer noticeable detail than just those vocal examples.

 

[SirPauly]

Tessellation is a buzz word and a scam.

Of course it is and very obvious by these shots:

http://forums.anandtech.com/showpost.php?p=32286552&postcount=554

 

[SirPauly]

it's not like that. You can 16-pixel tessellate every single surface in a game and the current video cards will cope with this. Subpixel tessellate only a few surfaces (and leave all the rest non tessellated) and the cards will struggle because a large amount of normal tessellation will not create a bottleneck while a small amount of hypertessellation will.

Crysis 2 is tesselated exactly the way you do it when you want to create bottlenecks and cripple the video cards.

The overtessellated objects are not even important and a gamer won't even look at those barriers or sidewalk borders let alone the water you can't see.

Maybe on some objects but what about the walls and such? A lot of the tessellation is on walls and its finer macro details may need the higher maximum to offer quality on them. Are there any examples of this kind of tessellation detail in any AMD sponsored game?

 

[SirPauly]

No one should be losing performance because of poor implementation of a feature though.

This youtube video shows a lot more objects and it looks like a lot of them they just broke the large polygons into thousands of tiny ones without actually adding any more information to the model. On many of them you can still see the way a ton of triangles just form 1 large flat triangle from the old model. There are clearly some spots where you can see the tessellation was used to deform and add depth though.

Also this John Carmack interview about tessellation was pretty interesting, and he points out tessellation is only really useful if you built from the ground up with it in mind.

I don't know about that; it is dark and moving -- how about taking static shots with and without tessellation and let's see. But, what the person did was find tessellation -- zoom in to show the amount of triangles and offer blanket views.

 

[Seero]

It was 6 pixels, not 16 and do you have the link? Or, can you point me to the post that has it. I would like to read his reasoning, thanks. Never mind.

Image taken from http://www.geeks3d.com/20101201/amd-graphics-blog-tessellation-for-all/

 

[Lonbjerg]

Image taken from http://www.geeks3d.com/20101201/amd-graphics-blog-tessellation-for-all/

What good is is that if you get artifacts on AMD a < 32x tessellation?

 

[AtenRa]

Here's a link with a breakdown of the performance hits. I've posted this link before.
http://www.behardware.com/articles/...ser-look-at-performance-and-tessellation.html

There's also a graph that shows what the Ultra performance hit is with AMD's tessellation slider set to 0.

http://www.behardware.com/articles/...ser-look-at-performance-and-tessellation.html

As during our tests AMD had no answer other than the drivers advised for use with Crysis 2 DirectX 11 were still the Catalyst 11.6s, we don't know.

Any FPS performance hit numbers with Cat 11.8 ??

AMD Cat 11.8
Performance Highlights

Improves performance up to 10% in Crysis 2 DirectX 11 version for both non-Anti-Aliasing, and application enabled Anti-Aliasing cases on the AMD Radeon™ HD 6000 and AMD Radeon™ HD 5000 Series
Improves performance up to 8% in Fear 3 DirectX 11 version with application enabled Anti-Aliasing on the
AMD Radeon™ HD 6000 and AMD Radeon™ HD 5000 Series
Improves performance up to 30% when AMD’s Morphological Anti-Aliasing (MLAA) is enabled through the
Catalyst™ Control Center on the AMD Radeon™ HD 6000 and AMD Radeon™ HD 5000 Series
Improves performance up to 20% in Call of Duty Black Operations for single GPU and Multi-GPU configurations on the AMD Radeon™ HD 6000 and AMD Radeon™ HD 5000 Series

So with Cat 11.8, HD6950 gets close to GTX560Ti, i believe that's OK and where it should be in Crysis 2 DX-11.
I don't believe anyone would like to see an HD6970 be close to GTX580 eee ??

 

[SirPauly]

So with future drivers, there seems to be performance gains, which was expected on AMD hardware. I don't understand why nVidia would sabotage themselves with all this wasted tessellation.

 

[96Firebird]

So with future drivers, there seems to be performance gains, which was expected on AMD hardware. I don't understand why nVidia would sabotage themselves with all this wasted tessellation.

They obviously didn't think AMD would figure out how to better their drivers for better performance. Those BASTARDS!