RDMA

Remote Direct Memory Access

RDMA (Remote Direct Memory Access) is commonly used alongside NVMe-oF. It allows direct memory access from one computer to another without CPU involvement, dramatically reducing CPU utilization.

Key Benefits

Provides significant advantages when computing and storage are disaggregated

Enables compute nodes to retrieve information from remote storage directly into memory without CPU intervention

Allows virtual machines on compute nodes to utilize more physical CPU resources

Implementation Considerations

Requires RDMA-compatible NICs and drivers

Hardware support must be verified before deployment

Storage Protocols

Several RDMA-based storage protocols exist:

NVMe-oF (NVMe over Fabrics)
SRP (Storage RDMA Protocol)
iSER (
iSCSI Extension for RDMA)

Mellanox has endorsed NVMe-oF as their development standard due to its superior performance

RDMA Protocol

iWARP

pplx-garden
perplexityai • Updated 2026 Jun 25 9:30

Despite the ConnectX-7 hardware dependency, it clearly analyzed the bottlenecks inherent in collective communication and presented practical engineering solutions to avoid them.

Notably, this technique increased the weight update speed by more than 100× in RL-based large language model training.

Weight Transfer for RL Post-Training in under 2 seconds

Ultra-fast cross-GPU model sync