NVIDIA Magnum IO GDRCopy is a low-latency GPU memory copy library based on GPUDirect RDMA technology, allowing the CPU to directly map and access GPU memory. It provides optimized copy APIs and is widely used in high-performance communication runtimes like UCX, OpenMPI, MVAPICH, and NVSHMEM. GDRCopy allows the CPU to directly access GPU memory through BAR mappings, enabling low latency copies between GPU and CPU memories.

Features

• Low Latency: Transfers data between host and device with very low latency, around 1 µs for host-to-device copies.
• High Bandwidth: Achieves high host-to-device memory copy bandwidth through write-combining or cached mappings.
• Optimized APIs: Provides optimized copy APIs for efficient memory transfers.
• Kernel Bypass: Avoids performance overheads of interrupt processing by providing a kernel bypass for send and receive queues.
• Multi-Architecture Support: Supports multiple CPU architectures, including x86, Power, and Arm.
• GPUDirect RDMA: Relies on GPUDirect RDMA, available on Tesla and Quadro GPUs.

Benefits

• Performance: Delivers maximum throughput at the lowest latency.
• Efficiency: Reduces latency and improves performance for small data sizes.
• Compatibility: Supports a wide range of CPU architectures and accelerators.
• Security: Provides hardware-based memory protection for secure applications.
• Integration: Easily integrates with high-performance communication runtimes.