Caffe

A deep learning framework from Berkeley AI Research (BAIR) built for speed, modularity, and high-performance computer vision.

Developed by Yangqing Jia at UC Berkeley, Caffe is an open-source framework optimized for convolutional neural networks (CNNs). It handles over 60 million images per day on a single NVIDIA K40 GPU (standard AlexNet speed). The architecture separates model configuration from implementation: users define layers and solvers using Google Protocol Buffer (Protobuf) files. This design allows for seamless switching between CPU and GPU training without rewriting code. While Caffe2 has since merged into PyTorch, the original Caffe remains a staple for large-scale industrial deployments and research projects requiring strict performance benchmarks.

https://caffe.berkeleyvision.org/