Accelerate deep learning & Habana Gaudi–based Amazon EC2 DL1 instances
summary
EC2 DL1 instances use Habana Gaudi accelerators (8 Gaudis/server in a full-mesh, 400 Gbps on‑die networking, 4 TB local NVMe, 32 GB HBM per chip) and SynapseAI software with Habana PyTorch/TF libraries. Demonstrated claims: typical 40–60% price‑performance vs leading GPUs; Mobileye reports >40% in production; one‑card BERT fine‑tune on SQuAD runs in 1m54s using bfloat16 and Hugging Face; scale uses MPI/torch DDP across eight Gaudis or tens–hundreds of nodes and works as shown.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 0:38Agenda
- 1:01Intel Habana Gaudi
- 2:16Price Performance
- 3:17Running a model
- 7:10Summary
- 7:50Models
- 10:50Real models
- 11:30GitHub page
- 15:38Hardware and Software
- 17:32Software Stacks
- 19:50Kubernetes
- 20:20Pretraining
- 21:18Mobileye
- 22:51Lidos
- 29:09Outro
related talks
31mRealize price-performance advantage for AI
31mRealize price-performance advantage for AI
50mThe democratization of tomorrow’s technology
36mAccelerating High-Performance Machine Learning at Scale i...
35mAccelerate your ML innovation with AWS ML infrastructure services
50mAI acceleration at the edge
28mWhat's new in Amazon EC2
27mThe convergence of machine learning and high performance computing