Hands-On GPU Computing with Python

Third generation

NVIDIA released GeForce 3 in 2001, and that gave programmers the ability to program on parts of the previously non-programmable pipeline we spoke of in the previous section. They could now send data along shaders with the ability to operate on it while being in the pipeline. The shader programs were small kernels, written in assembly-like shader languages. For the first time, a limited amount of programmability in the vertex-processing stage of the pipeline became possible. NVIDIA's competitor, ATI, also had its Radeon 8500 card available for consumers at the same time.

The GeForce 3 GPU had 57 million transistors, 64 MB of 128-bit DDR DRAM, with a 200 MHz core clock speed.

The Radeon 8500 GPU had 60 million transistors, ...