XBA and FPS

FS: Can you tell us more about this Xtreme Bandwidth Architecture (XBA)? Is XBA a marketing term used to describe the benefits of eDRAM?

From the XBA press release: "Bitboys plans to introduce the final product specifications, product names and target prices of the first XBA™ enabled products March 2000." Can we assume that these first XBA enabled products are all Glaze3D based?

Shane: XBA is really much more. eDRAM is of course something totally new in a 3D-graphics chip, but taking full advantage of it requires some intelligent new algorithms as we have many more times the data to move than in a traditional architecture.

Petri: Yes, the XBA technology was invented during the design of our current 3D-graphics architecture. XBA consists of the eight-texel/four-pixel rendering pipeline, extremely wide 512-bit memory bus and our memory management logic. The memory management logic works as a highway system tying together the embedded DRAM memory, the external SDRAM memory, AGP memory and all units, which want access to the memory. A lot of the XBA inventions are behind this memory system, which handles 768 bits of data (eDRAM+AGP+SDRAM) every clock cycle, resulting in the massive bandwidth of 12.5 GB/s.

This huge bandwidth enables us to do everything in true color; we are not really interested in any 16-bit performance and dithered images. The bandwidth also allows us to do full-scene Anti-Aliasing with real supersampling. Although rendering game graphics at 1024x768 true-color resolutions requires around 2.5 GB/s of memory bandwidth, doing this with anti-alias enabled requires 10 GB/s of memory bandwidth! To say this in another way, products equipped with only external SDR or DDR memory are not able to do full-scene anti-aliasing at realistic speeds. The XBA architecture based products are really the first chips capable of improving the image quality with Anti-Aliasing.

FS: Early reports have claimed that, in simulation, the Glaze3D can theoretically run Quake 3 at 200-300 frames per second. How does the real silicon perform compared to the simulations?

Considering CPU and AGP bandwidth limitations, will it actually be possible to hit 300fps in Quake 3?

Petri: The simulator is a cycle-accurate simulator, meaning that it works exactly like the final silicon. In fact, we use the simulator to generate all our test vectors for netlists and volume production.

At this time, there's no CPU that is capable of running Quake 3 at 300 fps, but Intel and AMD are making good progress and we expect great improvements with the next generation processors. As you only need 75 fps in reality, the rest of the rendering power is available for increased complexity and features like the above-mentioned Anti-Aliasing.