Glossary
What is NVMe?: rawcompute.in Glossary
NVMe (Non-Volatile Memory Express) is a storage access protocol designed specifically for flash-based SSDs, connecting directly via PCIe to deliver vastly lower latency and higher throughput than SATA or SAS interfaces.
NVMe is a host controller interface and storage protocol built from the ground up for NAND flash and next-generation non-volatile memory. Unlike SATA (limited to ~600 MB/s) or SAS (limited to ~2.4 GB/s per lane), NVMe drives connect directly to the CPU via PCIe lanes, eliminating the legacy storage controller bottleneck. A single PCIe Gen4 x4 NVMe drive delivers up to 7 GB/s sequential read and over 1 million random IOPS. PCIe Gen5 NVMe drives push this to 14 GB/s.
NVMe drives come in multiple form factors: M.2 (common in workstations), U.2 (2.5-inch hot-swap, common in servers), EDSFF (E1.S, E3.S. The emerging data-centre standard), and add-in cards (AIC). In servers, U.2 and EDSFF are preferred because they support hot-swap via front-accessible drive bays. The NVMe protocol also supports features like multi-path I/O, namespace management, and NVMe-oF (NVMe over Fabrics) for network-attached flash storage.
Why it matters when buying hardware
NVMe is the standard storage interface for all new server deployments. SATA SSDs are being phased out of data-centre use. When configuring a GPU server, NVMe drives are essential for fast dataset loading. Training pipelines can bottleneck on storage I/O if the data loader cannot feed the GPUs fast enough. Plan for at least 2-4 NVMe drives in RAID 0 or JBOD for local scratch storage. Ensure the server motherboard provides direct-attach NVMe slots or a PCIe-switched NVMe backplane for the number of drives you need.