Storage server

A storage server is a machine built to hold and serve data rather than compute on it. The budget goes to drive bays, controllers, and network bandwidth instead of cores or accelerators. The chassis gives up GPU slots for backplanes that address dozens of drives and for enough network ports to move what those drives can read. In a GPU cluster its job is feeding training data fast enough that the accelerators never sit waiting on I/O.

A 2U chassis commonly carries 24 drive bays. Dense 4U designs pass 60 bays, and top-loading JBODs reach past 100 in one box. Drives mix high-capacity SATA or SAS for bulk with NVMe for hot data. The network side runs 100 or 200 gigabit Ethernet, sometimes InfiniBand, so the host can deliver what the media is capable of. Parallel filesystems such as Lustre, IBM Storage Scale, and BeeGFS spread one namespace across many of these boxes, adding capacity and throughput by adding nodes.

Size to the read bandwidth the training pipeline needs, not just to total capacity. Work out the GB/s the GPUs will consume, then count the drives and network ports required to sustain it.

Sources

Last verified August 29, 2026.