Storage

What is NVMe?

Also known as: NVM Express
Definition

NVMe is a storage protocol that connects solid-state drives directly over PCIe, reaching roughly 500,000 random IOPS with sub-100-microsecond latency. Its deep parallel queue model is architecturally different from SATA, which inherits a single-queue design intended for mechanical disks.

Why NVMe matters

Its deep parallel queue model is architecturally different from SATA, which inherits a single-queue design intended for mechanical disks. For database-backed applications the random-IOPS gap is the difference users actually feel.

NVMe in practice

When you provision an NVMe instance, the first thing you look at is the queue depth of your storage workload. A SATA-backed box will max out around a few thousand IOPS under load; NVMe sustains roughly 500,000 random IOPS with sub-100-microsecond latency. Change your sysbench and fio tests to measure randread and randwrite, not sequential throughput, and watch the difference in database commits. If you optimise for a SAS disk's block size or ignore I/O scheduler settings, you're leaving most of the hardware's capability unused. Get it wrong and your latency climbs a hundredfold on workloads that actually touch disk.

What people get wrong about NVMe

People assume NVMe is just a faster SSD, so they pick a plan based on size in GB rather than IOPS. Correction: NVMe's difference is the parallel queue model, not just raw bandwidth. A huge SATA drive will still bottleneck at single-queue random I/O; a smaller NVMe drive will beat it for database workloads. Size matters, but for latency-sensitive work, IOPS is the metric that costs you if you ignore it.

NVMe — common questions

What random IOPS does NVMe achieve?

Roughly 500,000 random IOPS, according to the published definition. That is about a thousand times more than a typical SATA SSD. The sub-100-microsecond latency means that for small, random reads and writes, you are not waiting on disk mechanics or a single SATA queue — the controller and PCIe bus are the only path.

Do I need NVMe for a website?

Only if your workload is random-I/O heavy, like a database or a cache. Static file delivery or a low-traffic site won't notice the difference between NVMe and SATA. The gap shows up under sustained mixed workloads, where SATA's single queue saturates and latency spikes. If you run Postgres or Redis, NVMe will keep your tail latencies low.

More from storage

IOPS
IOPS measures how many individual read or write operations a storage device completes per second.
RAID 10
RAID 10 mirrors data across pairs of drives and stripes across the mirrors, giving both redundancy and high performance at the cost of half the raw capacity.
Snapshot
A snapshot is a point-in-time copy of a virtual machine disk, taken instantly and stored on the same infrastructure.