Contents

• 1. Overview
  • 1.1. How GPUDirect RDMA Works
  • 1.2. Standard DMA Transfer
  • 1.3. GPUDirect RDMA Transfers
• 2. Change log
  • 2.1. Changes in CUDA 13.0
    • 2.1.1. Deprecation Notice: NV-P2P API Support
    • 2.1.2. Device Attribute Updates
    • 2.1.3. Platform Notes: Orin (L4T and Embedded Linux)
    • 2.1.4. Known Issues
  • 2.2. Changes in CUDA 12.2
  • 2.3. Changes in CUDA 11.4
  • 2.4. Changes in CUDA 11.2
  • 2.5. Changes in CUDA 10.1
  • 2.6. Changes in CUDA 8.0
  • 2.7. Changes in CUDA 7.0
  • 2.8. Changes in CUDA 6.0
• 3. Design Considerations
  • 3.1. Lazy Unpinning Optimization
  • 3.2. Registration Cache
  • 3.3. Unpin Callback
  • 3.4. Supported Systems
  • 3.5. PCI BAR sizes
  • 3.6. Tokens Usage
  • 3.7. Synchronization and Memory Ordering
• 4. How to Perform Specific Tasks
  • 4.1. Displaying GPU BAR space
  • 4.2. Pinning GPU memory
  • 4.3. Unpinning GPU memory
  • 4.4. Handling the free callback
  • 4.5. Buffer ID Tag Check for A Registration Cache
  • 4.6. Linking a Kernel Module against nvidia.ko
  • 4.7. Using nvidia-peermem
• 5. References
  • 5.1. Basics of UVA CUDA Memory Management
  • 5.2. Userspace API
  • 5.3. Kernel API
  • 5.4. Porting to Tegra
    • 5.4.1. Changing the allocator
    • 5.4.2. Modification to Kernel API
    • 5.4.3. Other highlights
  • 5.5. Porting strategy for device drivers based on nv-p2p
• 6. Notices
  • 6.1. Notice
  • 6.2. OpenCL
  • 6.3. Trademarks