NVMe or SSD: does it matter for a website

For a small site the difference between NVMe and SSD is barely noticeable, because pages are served from memory rather than from disk. It shows up in shops and sites with a large database. At the same host NVMe costs only about $0.50–1 a month more, and sometimes less, so what usually decides it is not the disk type but its size and how the host has limited it.

Updated 28.10.2026 · Find a VPS for your needs →

If you run a blog, a business-card site or a landing page, visitors will not notice whether the server has NVMe or a regular SSD. A cached page is served from memory in milliseconds, and the disk barely takes part. The difference becomes visible when the site constantly reads from and writes to its database: in an online shop with thousands of products, in catalog search, during price list imports, in the WordPress admin with WooCommerce or in Magento.

So the rule is simple. If NVMe costs the same or slightly more, take NVMe. If you would have to cut memory to get it, take more memory on SSD instead: 2 GB on SSD is more useful for a site than 1 GB on NVMe, because a database that fits in memory hardly touches the disk.

How NVMe differs from SSD

Both are flash storage with no moving parts; the difference is how they connect to the computer. A regular SSD uses the SATA interface, which was originally designed for hard drives. Its ceiling is about 550 MB/s, and it handles a single queue of 32 commands. NVMe is connected directly to the PCIe bus: up to 3.5 GB/s on PCIe 3.0, up to 7 GB/s on PCIe 4.0, and there can be thousands of queues.

What matters for a site is not megabytes per second but the latency of small operations. A database reads and writes chunks of 4–16 KB in random places on the disk, and each visitor request may wait for dozens of such operations. On NVMe each operation takes less time, and under concurrent load the gap widens, because requests do not stand in a single queue.

There is an important caveat on a VPS. You do not get a physical disk but a virtual slice of shared storage, and the host almost always limits its speed so that one customer does not disturb the others. An "NVMe" disk with a tight limit can be slower than a good SSD without one. So the word in the plan is a reason to choose, not a guarantee, and the disk is worth testing on day one. I show how below.

How much more NVMe costs

Hardly any more. I compared plans with the same memory at the same host: I took the cheapest NVMe plan and the cheapest SSD plan and looked at the difference, using the accsly catalog data of 27 October 2026. There were between 88 and 236 such host pairs depending on the memory size. The catalog stores prices in roubles, so the dollar amounts are rounded.

With 2 GB of memory, NVMe at the median costs about half a dollar more per month, around 8%. With 4 and 8 GB, about a dollar more, or 7–10%. Over a year the extra cost is roughly $6–12. Meanwhile, at 40–50% of hosts with 2–8 GB plans, NVMe costs the same as SSD or less. It looks as if many hosts have already moved their main hardware to NVMe, and SSD remains in older product lines that are not always cheaper.

The catch is elsewhere: NVMe plans usually come with less space. With 2 GB of memory the median disk size is 30 GB on NVMe versus 40 GB on SSD, and with 4 GB it is 60 versus 75 GB. For a site with a large media library, or for keeping backups on the server itself, this can tip the choice towards SSD.

NVMe is now found in 40% of the plans in the catalog. Selection: NVMe VPS. If you need cheap space for archives and backups and speed does not matter, there are also hard-drive plans: HDD VPS.

When a site will feel NVMe

Take a WooCommerce shop with 20,000 products and a 3 GB database on a server with 4 GB of memory. The database does not fit entirely in memory, so catalog filters, search and checkout will read from the disk. Here NVMe speeds up the slowest pages, and that is worth the extra dollar. Plans for this kind of work: VPS for an online store.

The second case is heavy writing. By default, every committed change in MySQL or PostgreSQL is flushed to disk with a sync command, and the server waits until the disk confirms the write. A site with active comments, action logging, job queues or sessions stored in the database runs into exactly this. For such projects the VPS for databases selection is worth a look.

For a business-card site or a WordPress blog with page caching, the difference disappears into measurement noise. It is better to spend the money on memory or on a server closer to your visitors: a data centre in Frankfurt or London for a European audience, in New York or Virginia for the US East Coast. Selection: VPS for WordPress.

Which disk you actually got

Inside a virtual machine the type of the physical disk is usually hidden. The command

lsblk -d -o NAME,ROTA,SIZE,MODEL

shows a disk with a name like vda or sda and an empty model. A value of 0 in the ROTA column means the system considers the disk non-rotational, but the host can set this to anything. The name nvme0n1 appears when the host gives the machine a virtual NVMe device, yet even that says nothing about the real speed. Only a test tells the truth.

Testing the disk: fio and how to read the results

The most accurate tool is fio. On Ubuntu and Debian it installs with apt install fio. Run the tests on a fresh server before a live site is on it, because they load the disk. Each test creates a 1 GB file, so make sure there is space.

Test 1. Latency of a single operation. This is closest to what a site experiences:

fio --name=lat --filename=/root/fio.test --size=1G --rw=randread --bs=4k \
    --iodepth=1 --direct=1 --ioengine=libaio --runtime=30 --time_based

In the output, find the clat line in the read block and the average value avg in it, then the percentile table below and the 99.00th line. Units are shown in brackets: usec is microseconds, msec is milliseconds. On a good disk the average stays within 100–300 microseconds. If it is around a millisecond or higher and the 99th percentile goes into tens of milliseconds, the disk is overloaded or network-attached, and the database will be slow whatever the plan is called.

Test 2. How many operations the disk sustains under load.

fio --name=iops --filename=/root/fio.test --size=1G --rw=randrw --rwmixread=70 \
    --bs=4k --iodepth=32 --direct=1 --ioengine=libaio --runtime=60 --time_based

Look at IOPS= in the read and write lines. A few thousand operations per second is enough for a site or a small shop. If the plan is sold as NVMe but read plus write comes to less than 5,000–10,000, the host has capped the disk. A flat round number, for example exactly 10.0k on every run, also points to a limit.

Test 3. Synchronous write speed. This is how a database writes every committed change:

fio --name=sync --filename=/root/fio.test --size=256M --rw=write --bs=4k \
    --ioengine=sync --fsync=1 --runtime=30 --time_based

You need IOPS= in the write line: the number of confirmed writes per second. A thousand or more is good, and the database will write changes without waiting. If you get less than 200, every saved order or comment will wait for the disk.

The --direct=1 flag in the first two tests is essential. Without it fio measures the memory cache, and the result comes out several times too high. After the tests, delete the file: rm /root/fio.test.

A quick check with dd if fio is not available

The dd utility is present on every system. It is cruder but needs no installation. Sequential write:

dd if=/dev/zero of=/root/dd.test bs=1M count=1024 oflag=direct

The last line of the output shows the speed. A regular SATA SSD cannot deliver more than 550 MB/s, so figures above that indicate NVMe or fast storage. Less than 200 MB/s for an "NVMe" plan is a reason to ask the host about limits. But keep in mind that sequential writing hardly matters for a site: it is how large files are copied, not how requests are handled.

Synchronous writes in small blocks, closer to how a database works:

dd if=/dev/zero of=/root/dd.test bs=4k count=2000 oflag=dsync

Divide 2,000 by the number of seconds in the output, and you get how many writes per second the disk sustains. If dd finished in one or two seconds, that is a thousand writes per second or more, and all is well. If it took 10–20 seconds, each write takes 5–10 ms, and the database will feel it. After the check: rm /root/dd.test.

How not to misread the results

Run the tests two or three times at different times of day, for example in the afternoon and late in the evening. If the results jump by a factor of several, the disk is shared with busy neighbours, and the site will be slower at peak hours. Compare servers with the same commands: figures from different tools and different settings cannot be compared with each other.

Do this in the first days, while the refund period is still valid. Plans that offer one are in the money-back VPS selection. Testing the network and the processor is covered in the article how to check VPS speed, and the other plan parameters in how to choose a VPS.

FAQ

Do I need NVMe for a WordPress site?

For a blog or business-card site with page caching, the difference is barely noticeable. For a WooCommerce shop with thousands of products, NVMe speeds up filters, search and checkout, especially if the database does not fit in memory.

How much more does NVMe cost than SSD on a VPS?

At the same host with the same memory, the cheapest NVMe plan at the median costs about $0.50–1 more per month, according to the accsly catalog on 27 October 2026. At 40–50% of hosts with 2–8 GB plans, NVMe costs the same or less.

What matters more, NVMe or more memory?

For a site, memory matters more. A database that fits in RAM hardly touches the disk, so 2 GB on SSD is usually more useful than 1 GB on NVMe.

How do I test disk speed on a VPS?

Install fio and run a random read test with 4 KB blocks and the --direct=1 flag. An average latency of up to 300 microseconds is a good result. If fio is not available, dd with oflag=direct or oflag=dsync gives a rough estimate.

Why does the test show low speed on an NVMe plan?

The host almost always limits the speed of the virtual disk so that customers do not disturb each other, and neighbours may be loading the same storage. Repeat the test at different times, and if the result is consistently low, contact support or use the refund period.

When is an HDD enough?

For archives, backups and files that are rarely read. Do not keep a site or a database on a hard drive: random operations on an HDD are hundreds of times slower.

Plans on this topic

VPS for PostgreSQL Best VPS hosting 2026

More in the guide

Docker on a VPS: getting startedDocker lets you run a bot, a website, a database and monitoring on one server without them getting in each other's way, and moving to another VPS comes down to copying a couple of files and the data. Installation takes five minutes; the real trouble starts later: database ports exposed to the internet, logs growing to tens of gigabytes and data that vanished along with a container. Here is how to avoid all of that from the start.
How to move a website from shared hosting to a VPSThe move itself takes a couple of hours; the mistakes usually happen around it: a forgotten cron job, .htaccess rules that nginx does not read, an AAAA record with the old address, or mail that disappeared along with the IP change. Below is the order of steps with commands and checks for a PHP and MySQL site.
Server control panels: ISPmanager, FASTPANEL, aaPanelA panel turns a VPS into the kind of dashboard you know from shared hosting: sites, databases, mail and certificates are set up with a few clicks instead of editing config files. The three panels most often installed on Russian and budget international servers are ISPmanager, FASTPANEL and aaPanel. They differ in price, memory requirements and how much of the server they take over.
Windows VPS for 1CA rented server for 1C saves you from the office PC under a desk that nobody updates or backs up. But 1C on a VPS has its own quirks: the software licence is tied to the parameters of the virtual machine, users need both 1C and Windows licences, and core speed matters more than core count. Here is how to put together a 1C server for a small company.
Windows VPS: licensing, RDP and secure sign-inA Windows server costs more than Linux not only because of memory: the price includes a licence, and how it is arranged decides whether you are compliant and how many people can connect. An exposed RDP port is also the most common way such servers get hacked. Here are both topics.
When to move from shared hosting to a VPSA VPS does not make a site faster by itself: it gives you resources and freedom, and with them the job of running a server. Here is how to tell that hosting has become too small, when the problem can be solved without moving, and how to migrate a site so visitors notice nothing.
VPS backups: strategies and toolsOn a VPS, backups are your job, even if the plan says "backups included". We go through what exactly to back up, why a host snapshot is not enough, how to set up restic or BorgBackup in half an hour, and how to make sure you can actually restore from a backup.
First 10 steps after buying a VPS: securityBots start guessing passwords on a new server within minutes of it getting a public IP. Below are ten steps that close the most common holes in about an hour: commands for Ubuntu and Debian, the differences for AlmaLinux and Rocky Linux, and the checks that keep you from locking yourself out of your own server.
How to choose a country for your serverPut the server where your visitors are. For an audience across Europe, Germany, the Netherlands and Finland are convenient; for North America, the US East Coast. If you serve EU users, keeping the data in the EU makes GDPR simpler. Price hardly decides it: at the same host a server in Frankfurt, Amsterdam or New York usually costs the same, and only Asia and Australia are noticeably more expensive.
Hourly billing: when it pays offHourly billing pays off if you need the server for less than three weeks a month: for a test, a build, a one-off heavy job. At most hosts an hour costs exactly the monthly price divided by 730 hours, so over a full month you hardly overpay. The overpayment hides elsewhere: hourly plans are dearer on average, and for a server that runs all year, monthly billing or a year paid in advance usually comes out cheaper.
How much RAM and how many cores a website needsMost websites are fine with 2 GB of memory and one or two cores; a shop with thousands of products needs 4 GB and two cores. Memory matters more than cores: when it runs out, the site slows down or crashes, while an extra core without memory adds almost nothing. Below I show how to work out the memory for your site and how to see what it actually uses.
KVM or OpenVZ: which to chooseIn 2026 KVM is almost always the right pick: it costs a dollar or two more per month than OpenVZ, but everything runs on it, from Docker to Windows. OpenVZ makes sense in one narrow case, explained below, and at the end you will find the commands to check your own server in a minute.
What is a VPS, in plain EnglishA VPS is a slice of a powerful physical server that you rent and that behaves like a separate computer, with its own operating system, its own IP address and full control. Below is how it differs from shared hosting, who needs one, what it costs and what to check right after you buy it.
VPS for a Telegram bot: what the bot needs and how to pick a planA bot with a menu and a few commands runs fine on the cheapest server, while a bot with broadcasts, files and a database hits completely different limits. Here is what a bot really consumes, which Telegram limits shape the choice of server and how to avoid overpaying.
How to test VPS ping and speed before you payThe numbers in a plan do not tell you the latency to your users or how fast the disk really is. Most of this can be measured before you pay, the rest in the first hour after launch. Here are the tools and how to read the results.
Corporate network on a VPS: remote access to the officeA rented server can become the hub of a company's private network: remote employees and branch offices connect through it to the ERP, file server, CRM and Git. Here are the common layouts, protocols, server requirements and security basics.
How to choose a VPS: the 7 specs that matterTwo plans with the same "2 cores, 4 GB, 50 GB" can differ several times in speed and twice in price a year later. Here is what stands behind each line of a VPS plan and what to check before you pay.
VPS vs VDS: is there a difference and what actually mattersIn most markets VPS and VDS mean the same thing: a virtual server. The difference that really affects your work is hidden not in the name but in the virtualization type and whether resources are guaranteed.