More on TBR

Normalcy

Suppose the picture is of the two smiley faces we have below. (They aren't 3D, but for this example you will think they are.) Notice that the yellow face is partially over the blue face. Conventional rendering techniques will pre-render both of the faces and then use the z-buffer to determine which faces are visible.

So in the end, what you will see is the final image, two smiley faces. However, in a conventional polygon 3D accelerator, the graphics chip has to render both of the faces as if they had both been entirely visible. This is represented in the following picture:

More TB

If we toss tile-based rendering into the mix, the whole 3D rendering pipeline takes a major turn. First, the Kyro cuts up the given scene into little bite-sized pieces, in essence making tiles akin to a kitchen floor. Now before any rendering is done or textures passed, the on-chip z-buffer determines which parts of the image are going to be visible. This method makes more efficient use of the limited memory bandwidth by getting rid of unnecessary work, and most importantly, reducing all the needless amounts of texturing of scenes that will never be seen.

If you take a look at the picture below you will see what the Kyro was drawing. It never fully rendered the blue face because it wasn't necessary. While this example may be a bit simplistic, fire up any game and pause it. You will see many objects that are in front of one another, parts of cars, buildings, and other assorted items that you will never see, but were drawn many times over in the dark recesses of your video card. In the end this adds up to quite a bit of useless work. The ATI Radeon does something like this, they call it Hyper Z; although, they do not use a tile-based renderer.

It's a wonder we haven't seen other companies implement this method as of yet. If you look at the GeForce2, we can already see that memory bandwidth limitations are keeping that beast at bay. If you refer to our GeForce2 overclocking article you can see for yourself. Overclocking the GF2 core does almost nothing, but once the RAM is pushed up a bit, we can easily see measurable performance increases.