Originally Posted by Styves27

It should be noted that tessellation probably won't be used to such an extent in a real-world game.

Look at the dragon statue. It has far too many polygons than required. This is just to show potential, as TwoCables keeps having to mention. Real-world games will most likely use it for mild wall bumps rather than POM and perhaps characters. Tessellating to the point that they did in this video is more than overkill for games today.

I'd also like to point out the developer advantages to tessellation.

Traditionally, devs have to start with high polygon models, and then lower the detail so that they can work in-game. Then they create LOD models, even lower detailed models to swap with the regular model at distances, to lower polygon counts.

There are several problems with this:

1. It takes a lot of time. Creating the high polygon model takes a substantial development time. Then creating the real model based off the high polygon one as well as the LODs. This increases the development time (creating LODs isn't a simple click-done operation. In several cases, you end up with holes or oddities such as arms being flat and whatnot - I know from experience).

2. The swapping of each LOD is usually extremely noticeable. Of course, this depends on the view distance, but most of the time you can see the transition occur if you look for it. Some games (Killzone2) dissolve the current LOD into the next one, making the transition a little smoother, but it's still easily seen.

3. Memory. Each new LOD takes up memory. Normally, each LOD is embedded into the character model file itself. This means that the file ends up larger. Whether or not the system loads the entire set of models into memory or only the one currently in use, I'm not sure of, but I'm leaning towards the earlier. That would mean more memory being used for every model. Not good.

And here's where tessellation can help.

1. Devs can start straight off with a low-poly model, and simply tessellate the details onto it, as seen with the dragon or stairs in this video - that was the purpose of not making stairs, but a ramp instead. To show how effective tessellation can be to create detail. This decreases the development time significantly, because modeling a high-detail model will take more time then setting up some textures for tessellation. The same way modeling bricks would take more time than simply using POM.

2. Tessellation can be decreased in power over distance, as seen by the wireframe in the video. This means no swapping, only a smooth transition from full detail to low detail. You'd be hard pressed to see it occur (aside from a few things like the spikes on the dragon, which can be seen shrinking as you get further).

3. The only added memory is the memory created by the tessellation. There are no LOD models for every object, therefor a large portion of memory can be saved. Character models and objects can take up less HDD space and therefor devs can fit more onto a disk than they used to.

Performance shouldn't be an issue if they carefully select a proper level of tessellation. For example, tessellation in this video is over done for the purpose of effect and showing off the potential.

The last thing I'd like to note is the performance loss in Dirt 2 may be related to DX9 having a 1/4 screen-size HDR, SSAO and post-processing techniques (being already somewhat performance-lowering) while the DX11 one has full-screen HDR (much more accurate), SSAO and post-processing (blur, depth-of-field, that stuff).

Using a full-screen buffer for the shader effects in DX9 would undoubtedly lower performance beyond acceptable results. With DX11, it manages to stay above a playable FPS, so the difference is pretty large.

Just to show you what I mean, I created a Crysis map that has heavy lightbeam usage. Normally, these lightbeams are rendered on the HDR buffer, which is something like 1/4 or 1/8 the screen size. Using the console command "r_beams 1" the beams are moved from the HDR buffer to the normal buffer, resulting in a full-screen render. The beams end up much nicer and much more detailed, but performance is dropped by nearly 60%.

Couple that with full-screen SSAO and post-processing, like motion blur and depth of field (which are also rendered at around 1/4 the screen size) and you won't have a game anymore, just a screenshot.




Try to read it for goodness sake.

Originally Posted by BlankThis

Commencing operation Separation.
Primary objective: Save my keister

I will make a 50/50 solution of cat water and Vitamin E skin cream to produce a lubricating substance which I will rub on my behind. Then I will commence to lubricate the handle of my sacrificed tooth brush which I plan to use as a pry bar. I will use iTunes and my HD 555s to keep my mind off the task at hand. Once the majority of a cheek is separated I will use a section of a cardboard box from a mini fridge as a spatula.


I'm really nervous.

Originally Posted by TwoCables

Um. Whoa!

Originally Posted by flipmatthew

i use this benchmark, does it lag for u at 2xaa 16xaf all high dx11 tessalation when ur at the dragon??