Hyper-Z IIOne of the most innovative features of the original Radeon was the Hyper-Z memory bus. This feature allowed it to keep pace with rival cards with faster clock speeds by helping to escape the biggest bottleneck on any modern graphics card – the bus between the GPU and the card’s RAM. Hyper-Z allowed for advanced refinement and compression of the vertex data before it is stored or retrieved from card RAM, reducing the amount of data that needs to be sent over the bus. This is accomplished by several methods, including detecting unseen pixels in a frame and removing them and compression of the data used to sort polygons (known as z-buffer data).
The 8500 chipset features Hyper-Z II, a refined and turbocharged version of the original. Two interesting features allow this technology to outperform its predecessor. The first is a refined method for “culling” pixels in a scene which will not be rendered, and the second is “Fast Z-clear,” a method for clearing out the Z-buffer space instantly rather than having to purge it by writing zeroes to it as previously required (which wasted a clock cycle).
In a typical 3D scene, there will be areas of any given polygon that will be hidden by other polygons – for instance parts of a large polygon describing a floor surface will be hidden behind the models standing on the floor and the walls entering the scene from the sides. The game itself must send the size, shape and position of each polygon regardless of location; the intelligence needed to determine which pixels being processed will eventually make it to the screen when rendered should take place on the, or you are just burning CPU time that could go towards the actual game.
"Hierarchal Z," as ATI calls it, works by analyzing the image data in a grid pattern to determine whether a given pixel will be visible in the final rendering of a scene. The previous version of this feature could only detect 8 hidden pixels per clock cycle; now it can discard up to 64 pixels per clock cycle. This means up to 16 gigapixels per screen refresh can simply be thrown away and not rendered, saving on the amount of data that must be traded between the GPU and the video RAM.
When you add Fast Z-clear and Z-buffer compression to the mix, you get an overall 20% boost in performance for “free” – in other words, independent of the card clock speed or the speed of your computer’s CPU. The bottom line? This forward-looking approach gives the Radeon 8500 a bandwidth advantage in low-texture-memory situations, though such a scenario is tough to imagine with 64 MB of DDR RAM to play with.