Hercules 3D Prophet 4500 64MB - KYRO II (4/15)

The technology

As mentioned before, the Kyro's architecture departs from the classic rendering approach as employed by ATi's RADEON, 3dfx's Voodoo 3/5 or NVIDIA's GeForce line quite drastically. Instead of using the so-called "brute force" approach ("fillrate is King"), the Kyro goes about its rendering more intelligently and therefore efficiently. The simplified explanation of how the Kyro's tile-based rendering scheme works is as follows: The scene is divided into tiles of 32x64 pixels. The chip then renders the contents of each of these tiles in turn, one by one and writes them back to the video memory. There the tiles are recombined to form the completed scene. Thanks to internal caches on the chip, this happens very quickly and is unnoticeable for the user. Because of this trick, the Kyro II is not as heavily dependent upon and limited by memory bandwidth as "conventional" 3D chips, since it has to access the memory fewer times.

Yet this is only the beginning. The driver analyses the scene and ensures that invisible objects, say a door that is hidden by a person walking in front of it, are not rendered. Conventional chips would produce so-called "overdraw", since they render everything a game's 3D engine throws their way (Example: The visible scene in a game - What the 3D chip is really rendering (Transparent walls)). Besides needlessly increasing the polygoncount of a scene and necessitating more Z-Buffer checks, this also results in loading additional (nonetheless invisible!) textures into the card's memory, taking an additional toll on the already strained memory bandwidth.

A tile-based renderer like the Kyro and the Kyro II take a different approach. The 3D scene is divided into groups of polygons and sent to the chip, which manages these groups in a so-called display list. Now, before the polygons inside a tile are textured, the chip determines whether a given pixel is going to be visible in the finished scene or not. This process is called "deferred texturing" or "hidden surface removal (HSR)". This way, no textures need to be loaded for polygons that are going to be invisible to the viewer anyway (the door from our previous example) and the chip can forego the shading operations on them as well. 3dfx and NVIDIA are also working on minimizing or at least reducing overdraw. Just before the company went out of business, 3dfx released a beta driver that used HSR.

Table courtesy of PowerVR !!

NVIDIA's new GeForce3 is also capable of HSR, but comparing the performance increase with 3dfx's HSR drivers to the speed increase from the GF2 Ultra to the GF3, you can't help but wonder if there isn't still some work to be done...

Copyright: 09.04.2001 -   RIVA Station 2001 - Lars Weinand
No Copy without Permission!

Translation by Benjamin Kraft

URL of this Article: www.rivastation.com/3dprophet4500-64mb_e.htm - If you want to link to it, please use this URL! :-)