Voyager Part 2: Habana Architecture Deep Dive and Porting of TensorFlow and PyTorch Applications

Voyager is based on Intel’s Habana Lab AI processors and provides a unique opportunity to use AI focused hardware for exploring AI in science and engineering. Voyager features Habana’s Gaudi processors optimized for training, Goya processors optimized for inference, 100 GbE all-to-all connection within Gaudi nodes, 24 x 100GbE RDMA RoCE for scale-out across Gaudi nodes, and a Ceph file system.

This is the second of a two-part Voyager training session.

Agenda

• Porting applications to Habana Goya for inference 25 min
• Gaudi architecture for Deep Learning 25 mins
• Porting, optimizing and profiling PyTorch applications on Gaudi for training 40 mins
• Break 10 mins
• Porting, optimizing and profiling TensorFlow applications on Gaudi for training 40 mins