Skip to main content

6.2 Overview Of Voyager

Source repo: sdsc-summer-institute-2025 | Branch: main | Last synced: 2026-04-24 10:27:17.425 UTC

SDSC Summer Institute 2025​

Session 6.2 Overview of Voyager

Date: Friday, August 8, 2025

Summary: Voyager provides an innovative system architecture uniquely optimized for deep learning operations using well-established frameworks such as PyTorch and TensorFlow. Voyager comprises 42 training nodes of Supermicro X12 Habana Gaudi Training Servers; each training node contains 8 GAUDI HL-205 training processor cards which have 100 GbE non-blocking, all-to-all connections among the 8 cards within a node; the 42 Training nodes are connected via a high-performance, low latency 400 GbE switch interconnect. Voyager’s architecture has already shown highly scalable AI application performance in various areas such as LLMs (with billions of parameters such as for GPT2-XL and GPT3-XL), convolutional neural network-based image processing, and graph neural network based high-energy particle physics

Presented by: Amit Majumdar, Division Direction of Data-Enabled Scientific Computing

Reading and Presentations:​

  • Lecture material:
    • Presentation Slides: will be made available closer to the session
  • Source Code/Examples: N/A