Home › Guides › RDIMM, LRDIMM and UDIMM explained
Server memory explained: RDIMM vs LRDIMM vs UDIMM, ECC and ranks
Updated 13 September 2026
Desktop and laptop memory comes in one electrical flavour. Server memory comes in three, and choosing the wrong one is the most common reason a server refuses to post after an upgrade. This is what the letters mean, which machines take which, and the mixing rules that catch people out.
Unbuffered, registered, load-reduced
Every module presents its memory chips to the processor’s memory controller. The difference between the three kinds is how much help it gets doing so.
- UDIMM (unbuffered) — the chips talk to the controller directly. It is the cheapest and lowest-latency design, and the only kind an ordinary desktop takes. The ECC version (an ECC UDIMM, label code E) adds a ninth chip per rank for error correction and is what entry servers and workstations built on Xeon E3, Xeon E-2100/2200/2300, Ryzen Pro and some Core-based boards use.
- RDIMM (registered) — a register chip on the module buffers the address and command lines, so the controller drives one chip per module instead of eight or eighteen. That is what lets a server run twelve or twenty-four modules where a desktop manages four. All RDIMMs are ECC. Xeon E5, Xeon Scalable, Xeon W, EPYC and Threadripper Pro platforms use them.
- LRDIMM (load-reduced) — goes further and buffers the data lines as well, which is what makes 64GB and 128GB modules workable across every slot at full speed. Same platforms as RDIMM, but the two kinds cannot be mixed in one system: a board runs all-RDIMM or all-LRDIMM.
The board decides. A registered module will not post in an unbuffered board and vice versa, without exception. Look for the word in the manual, on the memory QVL, or on the label of what is fitted now (how to read it).
Which do I have?
| Platform | Takes |
|---|---|
| Ordinary desktop, any generation | UDIMM, non-ECC |
| Xeon E3 v3–v6, Xeon E-2xxx, Ryzen Pro, Core with ECC board | ECC UDIMM |
| Xeon E5 v3/v4 (Haswell/Broadwell-EP), DDR4 | RDIMM or LRDIMM |
| Xeon Scalable 1st–3rd gen (Skylake to Ice Lake) | RDIMM or LRDIMM |
| Xeon W-2xxx / W-3xxx, Threadripper Pro | RDIMM (LRDIMM on most) |
| EPYC 7001/7002/7003 | RDIMM or LRDIMM |
| Xeon Scalable 4th gen onwards, EPYC 9004 onwards | DDR5 RDIMM |
In vendor terms: Dell PowerEdge R630/R730 and later, HPE ProLiant Gen9 and later, Lenovo ThinkSystem and System x M5 onwards, and Supermicro X10/X11 boards are all DDR4 RDIMM (and LRDIMM) machines. Gen10 Plus, R650/R750 and X12 are also DDR4; the following generation moved to DDR5.
What 2Rx4 means, and why servers care
The rank is the number of independent 64-bit (72 with ECC) groups of chips on the module; the x4 or x8 is the width of each chip. A 16GB 2Rx4 RDIMM and a 16GB 2Rx8 RDIMM hold the same data and both work in nearly any DDR4 server, but they are not identical: x4 modules use more, narrower chips and support the stronger memory-error features (Intel calls it SDDC and ADDDC) on Xeon Scalable and E5. If the server documentation lists a rank for the memory feature you want, match it. Otherwise, either is fine, and mixing them across different channels is normal.
4Rx4 and 4DRx4 are quad-rank and appear on 32GB and 64GB LRDIMMs and on older 32GB RDIMMs. Quad-rank RDIMMs limit how many modules a channel can hold; LRDIMMs exist largely to avoid that.
Mixing rules
- Never mix RDIMM and LRDIMM. Never mix registered and unbuffered.
- Mixing brands is fine. Servers were shipped that way from the factory, and a Samsung M393A2K43 beside an SK hynix HMA82GR7 of the same speed and rank is a normal configuration.
- Mixing speeds works; everything runs at the slowest module’s speed, or at the speed the CPU allows for that many modules per channel, whichever is lower. A 2666 module in a 2400 server runs at 2400.
- Mixing capacities works on nearly all platforms if you populate channels sensibly (the larger modules in the first slot of each channel). Check the vendor’s population rules; Dell and HPE publish a table per server.
- Populate every channel before doubling up in any. A Xeon Scalable CPU has six channels: six 16GB modules outperform two 32GB ones by a wide margin, and cost about the same used.
Speed and the number of modules
A server rates its memory speed at one module per channel. Fill the second slot in a channel and many platforms drop a step: 2933 to 2666, or 2666 to 2400, particularly with dual-rank modules. This is normal and documented, and it is one reason to prefer fewer, larger modules once every channel already has one.
OEM part numbers
Dell, HPE, Lenovo and Cisco sell standard Samsung, SK hynix and Micron modules under their own
part numbers (Dell SNP…, HPE 8xxxxx-B21, Lenovo
4X77A… or 01AG…, Cisco UCS-M…). The
module underneath is the same part, and a server will run any module of the right type; the OEM
number matters only where a support contract insists on it. Where a module we sell carries an OEM
label, the listing quotes that number too.
Why used server memory is a sensible buy
Memory has no wear mechanism. Modules pulled from a decommissioned server were, by definition, running in a working machine the day it was switched off, and a module that passes a full test has no particular reason to fail later. The saving against a new module of the same part number is typically well over half, and on LRDIMMs and 32GB-plus capacities it is larger still. Everything here is tested at rated speed in a server before it is listed and again before it is packed.
Something here not match what you are seeing? Tell us — we would rather correct a guide than process a return.