What Is Storage Fabric?_
In GPU data-center deployments, the storage fabric provides the data pathway for loading training datasets, checkpoints, and model weights to and from the compute nodes. It typically uses a different protocol (e.g., NVMe over Fabrics or NFS over RDMA) and physical medium than the GPU-to-GPU NVLink or the scale-out InfiniBand/Ethernet fabric. This separation prevents storage traffic from interfering with GPU communication or job-to-job traffic.
Technical Details
The storage fabric is usually built with dedicated network adapters (e.g., ConnectX-7 or BlueField DPUs) and switches, often using 100/200/400 GbE or NDR InfiniBand, but it is logically and physically separate from the GPU compute fabric. Cabling is typically single-mode or multi-mode fiber with MPO connectors, and the topology is often a leaf-spine design to provide high bisection bandwidth. The fabric must support RDMA (e.g., RoCE v2 or InfiniBand RC) for low-latency data transfers to and from NVMe storage arrays.
How Leviathan Systems Works with Storage Fabric
In our field work, we run separate structured cabling bundles for the storage fabric, often using different color jackets or labels to avoid confusion with the compute fabric. We verify that storage NICs are in the correct PCIe slots and that switch ports are configured for the storage VLAN or partition before commissioning.
Related service
Network Testing & Commissioning →