Opening the GPU

Researchers were already using graphics chips for neural networks before general-purpose GPU tools became common. Programmable graphics hardware accepted data and small programs through interfaces built for rendering images.

In a 2004 implementation, researchers stored network values in texture data and used graphics programs to perform the calculations. The results came back through the same . It was neural-network arithmetic expressed in the forms the hardware already knew how to process.

announced in 2006 and released its first software tools in 2007.

CUDA did not invent GPU computing or neural-network training. It gave programmers a more direct way to send ordinary number work to the GPU's many workers.