NVMe Form Factors: U.2, U.3, M.2 and EDSFF Explained
NVMe drives come in several physical shapes, and unlike SATA and SAS — where a 2.5-inch drive is a 2.5-inch drive — the form factor determines what fits, how it is cooled, and how many you can install.
This guide covers the formats you will encounter in servers and how to work out which your chassis takes.
Why NVMe form factors multiplied
SATA and SAS inherited their physical formats from mechanical drives, which had to be 2.5 or 3.5 inches because that is what fits a spinning platter.
Flash has no such constraint. Once the platter disappeared, the shape of the drive became a design choice rather than a physical requirement, and different priorities produced different answers: density, serviceability, cooling, and the number of PCIe lanes each drive needs.
That is why NVMe has several formats where SATA has essentially one.
U.2 and U.3: the server standard
U.2 looks like a 2.5-inch drive and fits a 2.5-inch bay, but connects over PCIe rather than SAS or SATA. It is hot-swappable, front-accessible and serviceable exactly like a conventional drive.
That familiarity is why U.2 became the mainstream server NVMe format. It fits existing chassis layouts and existing operational habits.
U.3 is the successor and its main advantage is backplane flexibility. A U.3 backplane can accept NVMe, SAS and SATA drives in the same slots, which simplifies chassis design and lets an operator mix drive types without separate bay groups.
An important asymmetry: U.3 backplanes generally accept U.2 drives, but U.2 backplanes do not accept U.3 drives. Check which your chassis has before ordering, because the drives look identical.
M.2: compact, and often not what a server needs
M.2 is the small bare circuit board format familiar from laptops and desktops. In servers it appears mostly for boot volumes, frequently on a riser card or a dedicated internal mount, sometimes as a mirrored pair.
That role suits it well: a boot volume needs reliability rather than capacity, and putting it on an internal M.2 pair frees front bays for data drives.
Three limitations make M.2 unsuitable for primary server storage.
Not hot-swappable. M.2 drives are internal. Replacing one means opening the chassis and powering down.
Thermal constraints. A bare board in a confined space with limited airflow throttles under sustained load. Server-grade M.2 mounts include heatsinks for this reason, and it is still a constraint.
Lower capacity ceilings and endurance. The format limits how much flash fits, and many M.2 drives are specified for lighter duty than U.2 equivalents.
M.2 also comes in several physical lengths and two keying types, which are not interchangeable. Check both before ordering.
EDSFF: the density format
EDSFF — Enterprise and Datacenter Standard Form Factor — is a family of ruler-shaped and short formats designed specifically for dense flash deployment rather than inherited from mechanical drives.
The shapes allow more drives per rack unit, better airflow across the drives, and higher power delivery per drive than a 2.5-inch format permits.
You encounter EDSFF in high-density all-flash systems and in newer platforms designed around it. It is not something you retrofit — a chassis either supports it or does not, and the variants within the family differ in both length and thickness.
For most buyers maintaining existing servers, EDSFF is context rather than a decision.
PCIe lanes: the constraint behind everything
This is the part that makes NVMe different from every other storage decision, and it is where migrations get expensive unexpectedly.
SAS and SATA drives connect through a controller and a backplane. One controller can address many drives, and adding drives does not consume a scarce system resource.
NVMe drives connect over PCIe lanes, and a server has a finite number determined by its processors. Each drive consumes lanes — commonly four.
The arithmetic gets restrictive quickly. Populating twenty-four bays with NVMe requires far more lanes than most platforms provide directly, so dense NVMe systems use switching to share lanes among drives, at some cost in peak bandwidth per drive.
Two practical consequences.
You cannot simply swap SAS drives for NVMe in an existing chassis. The backplane, the lane budget and often the processors all have to support it. Our interface guide covers this.
NVMe drives compete with other PCIe devices. Network cards, HBAs, GPUs and accelerators all draw from the same budget. Adding NVMe storage can force choices elsewhere in the system.
Working out what your chassis takes
The reliable approach is the same as for drives generally.
Look at what is fitted. If the server has NVMe drives, read the part number and form factor off one. That is the format the backplane supports.
Check the bay markings. Many chassis label NVMe-capable bays distinctly, because on mixed backplanes only some bays are wired for PCIe.
Confirm U.2 or U.3. The drives look identical and the backplanes are not symmetric — U.3 accepts U.2, but not the reverse.
Check bay count against lane budget. A chassis with twenty-four bays may support NVMe in only some of them.
If you are unsure, send us the server model and generation and we will confirm which format and how many the platform supports before you order.
Do not forget endurance
Form factor determines what fits. It says nothing about whether the drive suits the workload.
An NVMe drive still has an endurance rating, and the same three classes apply — read-intensive, mixed-use and write-intensive. Buying the right shape with the wrong endurance class means either a drive that wears out early or one that cost considerably more than the workload justified.
Our SSD endurance guide covers how to calculate which class you need, and the SSD versus HDD guide covers whether the tier should be flash at all.
Sourcing
We source enterprise NVMe drives to order against exact part numbers across U.2, U.3 and M.2 formats. Send us your server model, the bays you intend to populate and what the drive will do, and we will confirm format, lane availability and endurance class before quoting.
The bulk quote page explains what to include, or email sarah.jane@techsellerusa.com.
Common questions
What is the difference between U.2 and U.3?
Both fit a 2.5-inch bay and connect over PCIe. U.3 backplanes can accept NVMe, SAS and SATA in the same slots, which simplifies chassis design. Importantly, U.3 backplanes generally accept U.2 drives but U.2 backplanes do not accept U.3 — and the drives look identical.
Can I use M.2 drives for primary server storage?
Generally not. M.2 is internal so it is not hot-swappable, throttles under sustained load in confined airflow, and has lower capacity and endurance ceilings. In servers it suits boot volumes, often as a mirrored pair on a riser, which frees front bays for data drives.
Why can I not just replace SAS drives with NVMe?
NVMe connects over PCIe lanes rather than through a storage controller, and a server has a finite number determined by its processors. The backplane, lane budget and often the processors all have to support it, so it is an architectural change rather than a drive swap.
Are all bays in an NVMe-capable server NVMe?
Frequently not. On mixed backplanes only some bays are wired for PCIe, and chassis usually label those distinctly. Check bay markings and the platform documentation before assuming a bay count.
What is EDSFF?
Enterprise and Datacenter Standard Form Factor: a family of ruler-shaped and short formats designed for dense flash rather than inherited from mechanical drives. It allows more drives per rack unit with better airflow, and it is not something you retrofit — a chassis supports it or does not.
Does NVMe compete with my network cards for resources?
Yes. Network cards, HBAs, GPUs and accelerators all draw from the same PCIe lane budget as NVMe drives, so adding NVMe storage can force choices elsewhere in the system.
Send us your server model and the bays you intend to populate and we will confirm format and lane availability before quoting.
