In our opinion, the current round of war between the red and green teams officially kicked off when ATI released their 5.11 catalyst driver- giving a nice boost in performance to the X1800XT, which, at its launch, didn’t really shake any ground up. The 5.11 Catalyst drivers basically allowed the X1800XT to surpass 7800GTX and nVidia fired back by releasing the 7800GTX-512 which pretty much wiped the slate clean. However, these 512 cards are still extremely expensive and hard to find- nVidia almost pulled a paper launch.

Now, yesterday we published an article on the X1800XT Crossfire and ended it by saying that you should wait a couple of days before deciding on a solution. Well, today, the red team fires back with the release of the X1900XTX (Lets hope that we dont get blocked by proxies by putting all those Xs together- seriously, whats up with that!) Not only is ATI fully prepared to have this card widely available as of today but they’re also releasing the CrossFire today as well. This is great leap from ATI considering most of their previous products have had to wait quite a bit before hitting store shelves.

Looking exactly like the X1800XT, the X1900XTX is based on 90nm process technology and packing a whopping 380 million transistors and is the first graphics card to feature 48 pixel shader processors- triple of what the X1800XT offered. Now this is a good thing considering almost all games released of late use Pixel Shaders. Here is a blurb from ATI regarding the types of Shaders and how it effects performance:

Shader instructions can be divided into two general classes: texture operations, which fetch data from memory, and arithmetic operations, which perform mathematical manipulation of data. While early shaders were divided roughly equally between these two types of instructions, more recent shaders tend to have a much higher proportion of arithmetic instructions. In the latest games, the average ratio of arithmetic to texture operations is approaching 5:1, and is projected to continue increasing in the next generation of game titles in 2006 and beyond.

Another observation is that in recent games, a majority of the pixels processed use bilinear filtering or point sampling from integer textures, or no texture lookups at all. These shader operations can be executed by each texture unit in one clock cycle or less. This balances against the remainder of the pixels which require trilinear filtering, anisotropic filtering, and/or floating point texture lookups, which require more than one clock cycle to execute.

One important difference between arithmetic and texture operations is that the latter are heavily dependent on external factors for performance, including graphics memory size and bandwidth. Adding more texture units can become fruitless without commensurate increases in these resources. Memory and bandwidth are in general more costly to scale than arithmetic operations, which are depend only on the number of processing units that can be fit into the GPU.

Procedural texturing is one important technique that can take advantage of this trend. Pixel shader programs can be used to generate textures mathematically based on a set of artist-controlled input parameters. This has the potential to dramatically reduce the amount of graphics memory and bandwidth required to store texture data. Alternatively, shader programs could be used to add variation and detail to existing textures, thus reducing the number of different texture maps that need to be stored in memory.