RDIMM vs LRDIMM vs UDIMM: which server RAM does your platform take?

Server memory comes in three module types that look nearly identical and do not interchange. Drop an RDIMM into a desktop board and it will click in like it belongs there, then refuse to POST. The reverse fails too. Before you buy, know which type your platform takes and how to confirm what a module is from its label, not from the listing title.

First, since people search for it: there is no “UDIMM vs DIMM” decision to make. Everything on this page is a DIMM, which is just the physical format (dual in-line memory module). UDIMM, RDIMM, and LRDIMM describe the circuitry on the module.

What each type actually is

UDIMM: unbuffered

On a UDIMM, the CPU’s memory controller talks directly to every DRAM chip on the module. Nothing sits in between. That keeps cost and latency down, which is why desktops use UDIMMs (laptops use SODIMMs, the same idea on a shorter board). The tradeoff is electrical load. The controller drives every chip directly, so a channel can only handle so many modules and ranks before signals degrade, which caps capacity. Entry-level servers and workstations, Xeon E boxes for example, typically take ECC UDIMMs, the same unbuffered design with extra DRAM to store error-correction bits.

RDIMM: registered

An RDIMM adds a register, a small chip in the middle of the module (the RCD, or register clock driver) that buffers the command, address, and clock signals between controller and DRAM. On DDR4, data still flows straight to the chips. Because the controller sees one load on the control lines instead of dozens, each channel supports more modules and more ranks reliably. Capacity and stability at full population, for roughly one clock cycle of added latency. Mainstream servers, Xeon Scalable and EPYC alike, run on RDIMMs, and server RDIMMs are ECC as a rule.

LRDIMM: load-reduced

An LRDIMM buffers everything. Alongside the RCD it adds data buffers, a row of nine small chips near the gold fingers that re-drive the data lines too. The controller now sees a single electrical load per module on every signal, which is what makes quad-rank and very large modules practical. DDR4 LRDIMMs are how 64GB and 128GB sticks became normal. You pay in price, power draw, and a little more latency than RDIMM.

The three types at a glance

	UDIMM	RDIMM	LRDIMM
What’s buffered	Nothing	Command, address, clock	Command, address, clock, and data
Where you find it	Desktops; entry servers (ECC versions)	Mainstream servers, DDR4 and DDR5	High-capacity DDR4 servers
Capacity per module	Lowest	High	Highest in the DDR4 era
Relative latency	Lowest	Slightly higher	Highest of the three
Marking letter	U (non-ECC) or E (ECC)	R	L
Mixes with the other types	No	No	No

How to read the label

Every server module carries two identifiers: a standardized marking and a manufacturer part number. Either one tells you the type in seconds.

The printed marking

Take a common label: 16GB 2Rx4 PC4-2666V-RB2.

• 16GB is capacity.
• 2Rx4 means two ranks of x4-organized DRAM chips. Rank and chip width affect population rules and advanced ECC modes.
• PC4 means DDR4. PC3 is DDR3, PC5 is DDR5.
• 2666 is the transfer rate in MT/s.
• V is a timing grade.
• R is the letter that answers this article’s question. R means RDIMM. U means non-ECC UDIMM, E means ECC UDIMM, L means LRDIMM, S means SODIMM.

That first letter of the final block settles most arguments. A module marked PC4-2400T-LB1 is an LRDIMM no matter what the listing said. PC4-3200AA-EB2 is an ECC UDIMM and will never boot in a Xeon Scalable box.

Manufacturer part numbers

Samsung encodes the type in the prefix. M393A2K40CB2-CTD breaks down as M393 (DDR4 RDIMM; M391 is DDR4 ECC UDIMM, M386 is DDR4 LRDIMM, M378 is plain desktop memory), A2K40CB2 (16GB, 2Rx4, die revision), and CTD (DDR4-2666; CRC is 2400, CVF is 2933, CWE is 3200). Samsung’s DDR5 server RDIMMs start with M321.

SK hynix puts the type letter mid-string. HMA84GR7AFR4N-VK is a 32GB 2Rx4 DDR4-2666 RDIMM; the R in GR7 means registered, and the 7 means a 72-bit ECC bus. An ECC UDIMM reads GU7 in the same position.

Micron puts it near the end. MTA36ASF4G72PZ-2G6 is a 32GB 2Rx4 DDR4-2666 RDIMM: PZ means registered, AZ means ECC unbuffered, LZ means load-reduced.

No label? Look at the board

A lone register chip in the center of the module means RDIMM. That chip plus a row of data buffers along the bottom edge means LRDIMM. Neither means unbuffered. Chip counts help too, at least on DDR4: a total divisible by nine (9, 18, 36) means ECC, while 8 or 16 chips means a desktop module. DDR5 changes the math, so treat divide-by-nine as a DDR4-and-older rule.

Compatibility rules

The platform dictates the type. No adapter or BIOS setting works around it.

• Mainstream server sockets (Xeon Scalable, EPYC) accept RDIMMs and, in the DDR4 generation, LRDIMMs. They do not accept UDIMMs of any kind.
• Platforms that spec ECC UDIMM take only that.
• Desktop boards take UDIMMs. A few accept ECC UDIMMs; none take registered memory.
• RDIMM and LRDIMM cannot be mixed in one system even when the board supports both. Pick a type for the whole box.
• Within DDR4, all three types share the same 288-pin connector and key position, so a wrong-type module inserts without resistance and simply fails memory training. DDR4 and DDR5 are keyed differently and cannot cross generations.

When in doubt, the server’s population guide or memory QVL lists exactly what each configuration supports.

When LRDIMM is worth it

LRDIMM earns its premium in one situation: maximum capacity per socket. If the build fills every slot with the largest modules the platform allows, load reduction is what keeps fully populated channels running at speed. If the target capacity is reachable with ordinary dual-rank RDIMMs, those are cheaper and marginally faster. Late in the DDR4 era, 64GB dual-rank RDIMMs built on 16Gb dies took over territory that used to belong to LRDIMMs, so the price relationship between the two keeps shifting on the secondary market. Quote both and compare.

What changes with DDR5

DDR5 rewrites some of these rules. Power management moved from the motherboard onto the module (every DDR5 DIMM carries its own PMIC), and each module is split into two independent subchannels. The practical outcome for servers: DDR5 is RDIMM-dominant. The registered ECC module is the default and, so far, nearly the only option on mainstream server platforms, with ECC UDIMMs covering entry-level machines.

There is no DDR5 LRDIMM in the DDR4 sense. High capacities come from RDIMMs built with stacked (3DS) DRAM, and a newer class of multiplexed-rank modules is emerging to push bandwidth by running two ranks in parallel: MRDIMM, plus the Intel and SK hynix MCR DIMM that preceded it. Expect those in volume on the secondary market in a few years, not today.

One caution that comes up constantly: DDR5’s on-die ECC is not server ECC. Every DDR5 chip corrects certain errors internally, but that protects nothing on the bus and does not turn a desktop UDIMM into ECC memory. A listing that calls plain DDR5 “ECC RAM” is mislabeled, whether the seller realizes it or not; check the part number before relying on it.

Notes from the trade

Mixed lots get discounted. A pallet described as “1,000 x 32GB DDR4” that turns out to blend RDIMMs, LRDIMMs, speeds, and ranks cannot be resold as-is. Someone has to pull every module, read every label, test, and re-lot by full marking, and that labor comes out of the offer. Sellers who sort by complete part number before asking for quotes get better numbers, consistently. If you have memory to move, send a list with part numbers and label photos rather than a summary line.

The label outranks the description. We quote from markings and part numbers because listing titles are wrong often enough to matter; “server RAM 32GB” could be any of three types at any of five speeds. Buying from anyone, ask for label photos before money moves. Buying from us, part numbers get you an exact answer fast. Our server memory stock rotates, quotes come from what is on the shelf, and you can request a quote without opening an account.

Most of this comes down to two checks: the type letter in the marking and the platform’s population guide. The server memory purchases that go sideways usually skipped one of them.