NVIDIA DGX Station Development Guide

NVIDIA DGX Station is part of a new class of computers designed to build and run AI. It is the first system built with the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip and up to 748 GB of large coherent memory — delivering substantial compute performance for developing and running large-scale AI training and inferencing workloads at your desktop. Combining advanced system capabilities with the NVIDIA AI Software Stack, NVIDIA DGX Stations are purpose-built for teams that need a high-performance desktop AI development platform.

• Introduction to DGX Station
  • DGX Station - Designed to Build and Run AI
  • Built on NVIDIA Grace Blackwell Architecture
  • Up to 748 GB of Coherent Memory for up to 1T Parameter Models
  • Bringing Blackwell Architecture from the Cloud to Your Desk
• System Overview
  • Hardware
  • Software
• ARM CPU Architecture
• SIMD
• Porting Existing Apps
  • Dependencies & Tools
  • DGX Station Software Stack
  • Setup and Build Systems
  • Compilation on DGX Station
  • CPU Instruction Extensions
  • ARM Memory Ordering
  • Page Size
• Optimization for DGX Station
  • General Practices
  • Compiler Configuration
  • gcc and clang
  • .NET Applications
  • Taking Advantage of UMA
  • Coherent Memory Management
  • Profiling
  • Links
• Mixed Coherency Environment
  • Coherent Memory
  • Non-Coherent Memory
  • CUDA Unified Virtual Memory
  • Querying for Coherency-Mode Support
  • UVM Pitfalls
  • Current Limitations of the Mixed Coherency Environment
  • Future Updates to the Mixed Coherency Environment
  • Vulkan-CUDA Interoperability
• Feedback