TensorFlow
Google open-sourced its own system for training models: computation is described as a graph and can run on a processor, a graphics accelerator or a cluster.
Why it matters
An industrial-grade tool became free, and the barrier to deep learning fell to knowing Python.
Before this every laboratory had its own half-written tooling and reproducing someone else's result was hard. TensorFlow gave a shared foundation and with it a shared way of describing models. A year later PyTorch appeared with a different execution philosophy, and the rivalry between the two shaped the field's tooling for a decade.