Rage 128 did a good job of doing what DirectX 6.0 expected of it a few years ago - triangle setup in hardware for example. It wasn't a Voodoo in that it didn't change anyone's expectations of 3D graphics, but it met the requirements of ATI's OEMs admirably delivering a minimum feature set for a pretty good gaming experience. It was designed for the masses, not gamers, and in that regard it did its job.

You get to Radeon, and ATI is now starting to think about the competition - Nvidia's making waves with GeForce and hardware T&L. Radeon hits the DirectX 7.0 requirements, and introduces its own hardware T&L, a multi-pipeline renderer, and HyperZ.

However, this is still a fixed function graphics pipeline. 3D graphics hardware, up until the advent of DirectX 8.0 this year, remained limited to general purpose 3D commands. DirectX 8.0 has turned graphics chips into fully programmable devices, and that's good news for ATI.

If you compare the architecture of the Rage 128 to the Radeon you'll probably be able to guess where the next generation of changes are going to take place in ATI's architecture. Fundamentally, we can break down the next generation of graphics chips into a geometry processor (Vertex Shader), rasterizer (Pixel Shader), and memory controllers that optimize vertex caches, z-buffering, texture caching, and interfaces to the system bus. Yes, we will see more rendering pipelines, but if we expect greater memory bandwidth, and better performance from our geometry engines then we should expect pixel operations to increase. So, the split between Charisma Engine and Pixel Tapestry is going to be basically between a number crunching geometry processor front-end, and a parallelized back-end rasterizer. ATI can improve the performance and the number of operations in the Charisma Engine. ATI will increase the number of rendering pipes in its next generation.

How ATI handles memory in its next generation is another issue. That's not so clear. Do they stick to the existing model, or do something drastic in hardware? Will it be some sort of new cache structure, or algorithm for the Vertex Cache? Will ATI add some more brute force to its architecture? That's going to be the real issue, or is the company going to try and finesse its way out of bandwidth limitations?

I remember the launch of the Voodoo 5 in Las Vegas a few years ago. 3dfx was emphasizing all the rendering effects they had implemented, anti-aliasing, motion blur etc. At the same time, Nvidia was talking up its GeForce plans and hardware T&L. That was a tough one to call then for some people because, the focus was on what was going on in the 3D pipeline - 3dfx was using finesse, and Nvidia had brute force. Now, if ATI goes up against Nvidia, sure they can duke it out over their feature set, but in reality, I believe the hardware they put out has to have an element of brute force.

ATI's going to either have to do something with embedding memory in its next generation graphics, or finding ways to improve its memory controller and interface. Can they do that? I don't know if ATI can be that daring. The company has been a little conservative in the past. Can they be bold, and aggressive? Even if they don't come up with a genuinely superior solution, ATI certainly needs to add some dimension to their architecture that sets it apart from Nvidia. Nvidia did that with hardware T&L, and that was a watershed for them. I think Nvidia will do it again, but ATI hasn't done anything similar in the 3D world.