What Is IOPS?

IOPS, short for input/output operations per second, is the number of individual read or write requests a storage device completes in one second. Each request might fetch a 4 KB database page, write a log entry or read part of a file. IOPS counts the requests, not the bytes, so it describes how well a drive handles many small jobs rather than one large transfer.

That makes it the figure that matters for operating systems, databases, virtual machines and any application that touches lots of small files. For streaming large files, such as video recording or copying images, throughput in MB/s matters more.

An IOPS figure means little on its own. It always depends on the block size, whether access is random or sequential, the mix of reads and writes, and how many requests are queued at once. Two datasheets can quote very different numbers for the same drive simply by testing it differently.

Three numbers describe storage performance, and they are linked. Throughput equals IOPS multiplied by block size: 100,000 IOPS at 4 KB is about 400 MB/s, while the same throughput in 1 MB blocks needs only 400 IOPS. Latency is the time one request takes. With one request in flight, IOPS is roughly one divided by latency, so a 100 microsecond response gives about 10,000 IOPS.

Drives reach their headline IOPS by working on many requests at once. Queue depth is the number of requests waiting, and by Little’s law IOPS is roughly queue depth divided by latency. A datasheet figure measured at a queue depth of 32 or more may be far above what a single-threaded application sees at a queue depth of one or two.

The figures below are typical 4 KB random read results. Random writes are usually lower, and real systems add file system and RAID overhead.

The jump between SATA and NVMe comes from the interface: NVMe supports thousands of parallel queues where SATA has one. Each PCIe generation doubles the bandwidth per lane, and our FAQ on how PCIe 4.0 and 5.0 compare covers what that means at board level.

Datasheet IOPS are best-case numbers. The Storage Networking Industry Association (SNIA) Solid State Storage Performance Test Specification exists because SSD performance changes with use, and a few checks help compare drives fairly.

Most industrial PCs need fewer IOPS than a desktop. A machine controller, gateway or HMI reads its application at start-up, then writes small logs, well within the reach of any SSD. Endurance and temperature rating matter more there than IOPS.

Demand rises with concurrency. A local database, several virtual machines or an AI pipeline writing detections and metadata for many cameras all generate small random I/O, and that is where NVMe pays off. Industrial NVRs are the opposite case: recording is mostly large sequential writes, so throughput and endurance set the drive choice. Data-heavy nodes such as edge AI servers use multiple NVMe drives to keep the GPU fed.

To size storage, measure the workload with a tool such as fio on Linux or DiskSpd on Windows, note the block size, mix and queue depth, then add headroom for growth and RAID overhead.

We fit SATA or NVMe storage to every system we build as standard, sized from the workload’s block size, read/write mix and endurance needs, and validated in the system before it ships.