NVMe vs SSD for VPS: What is the real difference in practice?

Hosting providers highlight NVMe as one of their key selling points, and customers buy it because NVMe is faster than SSD. The truth, however, is more complex: the difference between NVMe and SATA SSD is real and critical in certain scenarios, but with VPS other factors come into play that can significantly diminish or amplify this difference. Let's look at the facts.

Technical difference: How NVMe and SATA SSDs work

Classic SSD drives, which replaced HDDs, typically use the interface SATA III - originally designed for mechanical disks. Maximum SATA III throughput is 600 MB/s. Access latency is typically 50-100 microseconds.

NVMe (Non-Volatile Memory Express) is an interface designed specifically for flash memory. It connects via the PCIe bus, which offers significantly higher throughput. Top-tier NVMe drives achieve throughputs of 3 500-7 000 MB/s (sequential read) and latencies under 20 microseconds.

Parameter	SATA SSD	NVMe (PCIe 3.0)	NVMe (PCIe 4.0)
Max. Sequential read	~550 MB/s	3 000-3 500 MB/s	6 000-7 000 MB/s
Max. Sequential entry	~520 MB/s	2 500-3 000 MB/s	5 000-6 500 MB/s
Random read 4K IOPS	~97 000	400 000-600 000	800 000-1 000 000+
Access latency	50-100 μs	10-20 μs	5-15 μs

On paper, NVMe is 5-10× faster in sequential operations and 4-10× faster in random IOPS. These are dramatic figures, but for VPS users the key question is different: what storage will I actually receive?

VPS Reality: Virtualisation changes the equation

When you order a VPS with NVMe, you are purchasing access to virtualised storage. Your disk is not a dedicated physical NVMe drive; it is a volume on a shared storage system that may (but does not have to) be based on NVMe hardware.

Storage architecture for VPS

Hosting companies usually use one of these approaches:

• Local NVMe (dedicated storage per node) - Each physical server has its own NVMe drives and the VPS on this server runs on these drives. The fastest option, but without redundancy in case a drive fails.
• Distributed storage (Ceph, ZFS replication) - data are replicated across multiple servers. It uses NVMe disks, but performance is lower due to network replication. Typically 200-500 MB/s effective throughput per VPS even on NVMe.
• SAN/iSCSI storage - centralised storage server via network. It may feature NVMe under the hood, but VPS performance is limited by the network layer.

IOPS limits per VPS

Even if the physical disk under your VPS is NVMe with 500 000 IOPS, your VPS typically has an IOPS limit set, so that one VPS cannot saturate the storage for other customers on the same host server. Typical limits range from 1 000 to 20 000 IOPS per VPS, regardless of whether SATA SSD or NVMe is used underneath.

This does not mean that NVMe is irrelevant. It matters. Lower NVMe latency will also be noticeable even with limited IOPS. However, you are unlikely to see the maximum figures from marketing materials (600 000 IOPS).

When NVMe Really Matters

Database workloads

This is the area where the difference between NVMe and SATA SSDs is most pronounced and measurable. MySQL, PostgreSQL, MongoDB and other databases generate intensive random I/O, especially during complex queries, JOIN operations, or when working with indexes that do not fit into RAM.

For a database server with 8 GB RAM and a significantly larger database, NVMe storage with lower latency will be measurably faster. A benchmark of the production workload (not a synthetic test) on PostgreSQL with typical e-shop queries typically shows a 20-40% performance improvement when switching from SATA SSD to local NVMe.

E-shops and applications with intensive I/O

WooCommerce, Magento, Shoptet, applications with thousands of products, categories and customers. Each page load generates dozens of database queries. When sharing a database server with multiple websites or during higher traffic, I/O latency becomes a bottleneck before CPU or RAM.

Containers and microservices

Docker and Kubernetes workloads involve frequent reading and writing of container images, logs, and configuration files. Fast storage reduces container startup time and deployment time.

Development servers with compilation workloads

Compiling large projects (Go, Rust, C++, Node.js bundle) generates intensive I/O. A large Go project compiles significantly faster on NVMe than on a SATA SSD: the difference can be up to 50% depending on the project.

When NVMe Doesn't Matter

Static websites and WordPress with caching

If your website primarily serves static content or has page caching configured (Redis, Varnish, or a WordPress cache plugin), the vast majority of requests serve content from memory (RAM) rather than from disk. In this case, disk speed is irrelevant.

Mail servers

A standard email server with a few thousand accounts is not I/O-bound; it is limited by the network and CPU for spam filtering. A SATA SSD is entirely sufficient for most email servers.

VPN servers, load balancers, proxies

These workloads are network-bound. Their performance is affected by bandwidth and CPU, not disk speed.

Backup servers

Backup workloads are sequential and network-bound. NVMe will deliver faster local write speeds, but the effective backup rate will be limited by network throughput.

How to spot fake NVMe marketing

Unfortunately, "NVMe" in a hosting company's offer is not a standardised term: and some providers use it creatively:

• NVMe-backed storage - This may indicate that NVMe drives are located at one end of the distributed storage. The actual performance of your VPS could be 200 MB/s sequentially.
• Missing specific IOPS limit - if the provider does not specify guaranteed IOPS for the VPS, it is likely that they either share performance without limits or the limits are low.
• Price too low for the declared performance - Real local NVMe with dedicated IOPS for each VPS is more expensive than SATA SSD. If the price looks too good, it probably is.

How to verify actual performance after purchasing an air-to-air heat pump:

# Random IOPS Test (4K Read/Write)
fio --name=random-rw --ioengine=libaio --iodepth=16 \
  --rw=randrw --bs=4k --direct=1 --size=1G --numjobs=2 \
  --runtime=30 --time_based --group_reporting

# Simple Sequential Read Test
hdparm -Tt /dev/sda

Then compare the results with what was offered in the proposal. If you receive 50 000 IOPS instead of the promised 300 000, the situation is clear.

Exascale.cz operates VPS exclusively on local NVMe disks with Proxmox KVM virtualisation, without a distributed storage system that would dilute performance. Each VPS accesses the NVMe disks directly on the host server.