As we’ve repeatedly pointed out in our earlier reports, the leading GPU developers ATI and Nvidia take significantly diverse approaches to designing new GPUs. ATI Technologies, currently the graphics division of AMD, has long been putting an emphasis on the speed of math1ematic computations and considering other parameters such as texture sampling and rasterization performance as of secondary importance. This developer had been sticking to the 3:1 concept until the arrival of unified-architecture cores as this ratio of math1ematical instructions to texture operations was thought optimal.
Of course, the ratio of 3:1 was not a developer’s whim but the result of the analysis of many factors, particularly that a lot of gaming projects were developed for several platforms at once. Game consoles had limited amounts of graphics memory and game developers had to use low-quality textures, putting a stress on visual effects created by means of math1ematics-heavy shaders. The product of this development approach was the ATI R580 core that contained 48 pixel shader processors and only 16 TMUs, yet became the best GPU with a non-unified architecture. R580-based graphics cards, ATI Radeon X1900 XTX and later ATI Radeon X1950 XTX, deservedly held the title of the highest-performance solutions with DirectX 9.0c support.
It is also then that the drawbacks of the 3-to-1 concept became apparent because high display resolutions and full-screen antialiasing were widely used in the PC world. Moreover, shaders used to imitate materials contained a lot of texture sampling operations anyway, which put a load on the TMUs and memory subsystem. Nvidia responded to ATI’s success with the release of the G71 core that differed from the previous G70 in thinner tech process and higher clock rates. Nvidia’s combination of 24 pixel shader processors per 24 TMUs proved to be no less, and sometimes even more, efficient than ATI Technologies’ 48/16 formula.
The computing capacity of GPUs has increased considerably with the arrival of the new generation of graphics cores with unified architecture. Execution processors now have to process three types of shaders, and it is impossible to tell the ratio of computing units to texture sampling and filtering units. The performance of streamed processors can only be estimated by indirect evidence now.
Today we’ll discuss the GeForce 8800 GTS graphics card, currently the most math1ematically advanced solution from Nvidia.