Protocol
GPUDirect RDMA
NIC DMA path that reads and writes GPU memory directly, bypassing host bounce buffers for RDMA verbs. Requires peer-to-peer BAR mapping, IOMMU policy, and registration of GPU memory regions. It underpins high-performance multi-node GPU training and inference fabrics.
Domains
Layers
Tags
gpuzero-copyp2pverbs
Sources
- NVIDIA GPUDirect RDMA documentation
Related protocols
RDMA Verbs
Lab readyKernel-bypass networking model: queue pairs, memory regions, and work requests move data between registered buffers with minimal CPU. Com…
dma-buf
Lab readyLinux buffer-sharing primitive for zero-copy exchange across drivers such as GPU, NIC, and media. Exporters and importers negotiate acces…
NVLink
Lab readyNVIDIA high-bandwidth GPU interconnect for peer-to-peer and coherent-ish memory traffic inside a node or NVSwitch domain. Generations dif…
GPUDirect Storage
Lab readyGPUDirect Storage (GDS) provides a direct DMA path between NVMe and GPU memory, skipping host bounce buffers. Filesystem and driver suppo…