What is RAID?: rawcompute.in Glossary

RAID groups multiple physical storage drives into one or more logical volumes using either a hardware RAID controller or software-based RAID. Common RAID levels include RAID 0 (striping for performance, no redundancy), RAID 1 (mirroring for redundancy), RAID 5 (striping with distributed parity, tolerating one drive failure), RAID 6 (double parity, tolerating two failures), and RAID 10 (mirrored stripes, combining performance and redundancy). Each level represents a different trade-off between usable capacity, read/write performance, and fault tolerance.

In data-centre servers, hardware RAID controllers from Broadcom (MegaRAID) or Microchip (Adaptec) offload parity calculations from the CPU and include battery-backed cache to protect in-flight writes during power loss. For NVMe-only configurations, many operators use software RAID (mdadm on Linux, ZFS, or Storage Spaces) since hardware NVMe RAID controllers are less common and NVMe drives are fast enough that CPU-based parity computation is not a bottleneck.

Why it matters when buying hardware

Choose your RAID level based on your workload’s tolerance for downtime and data loss. For boot drives, RAID 1 is standard. For large storage arrays serving AI training datasets, RAID 6 or RAID 10 with hot-swap drives provides a good balance of capacity, performance, and protection. For NVMe scratch storage in GPU servers (used for temporary training data), JBOD or RAID 0 may be acceptable if data is replicated elsewhere. Rawcompute.in configures RAID as part of server provisioning. Specify your requirements at order time.