Availability · 2007
NVIDIA opened a software platform for general-purpose computation on graphics processors to developers.
Access to parallel computation on ordinary graphics cards made training large models practical outside specialised computing centres.
The year is marked approximate: the beta and the first stable release both happened during 2007, and the archived page snapshot cited here is from 2008.
NVIDIA (archived) · Published 2008
The AlexNet paper describes training on GPUs using this platform.
ImageNet ClassificationBoth records are about the computing platform models are trained on.
The class of device CUDA opened to general-purpose computation eight years later.
The other half of what made training large models affordable: renting computation instead of buying it.
The computation runs on graphics processors through CUDA.
Large-scale deep unsupervised learning using graphics processorsInstead of a repurposed graphics chip, a chip designed for neural networks.