RDIMM vs LRDIMM vs MRDIMM: The DDR5 Server Memory Guide

For most DDR5 AI and general-purpose servers, RDIMM is the default: it is the registered, ECC module every Sapphire Rapids, Emerald Rapids, Xeon 6, and EPYC Genoa/Turin platform is qualified for, spanning 16GB to 256GB. Reach for 3DS LRDIMM when you need maximum capacity per slot — the 128GB-and-up high-density tier — and accept a small latency cost for the added data buffering. Choose MRDIMM 8800 when memory bandwidth, not capacity, is the bottleneck: it multiplexes two ranks to push roughly 37.5% more bandwidth than DDR5-6400 RDIMM on supported Xeon 6 hosts. All three are ECC modules built on dies from Samsung, Micron, and SK Hynix, then assembled by the same makers or by Taiwan and China module houses. The practical catch in 2026 is supply, not selection: DRAM is in an acute, AI-driven shortage, with prices up roughly 480–500% year-on-year and lead times of 30–40+ weeks (as of 2026-08). High-density, high-speed, and all MRDIMM parts move on allocation and RFQ rather than stable list pricing. This guide covers how each type works, the ranks and capacities involved, and how to spec the right module.

How each module type works

A DDR5 server DIMM carries side-band ECC and, on registered parts, a register chip that buffers command and address signals so the CPU can drive more modules cleanly. That is a plain RDIMM — the workhorse. Nearly every mainstream capacity from 16GB to 96GB, and octal-rank 256GB parts, ships in this form.

A 3DS LRDIMM (Load-Reduced, 3D-Stacked) goes further: it buffers the data lines too and stacks multiple DRAM dies inside each package. That reduces the electrical load the memory controller sees, which is what makes very high densities — the 128GB-class 4Rx4 modules — practical without wrecking signal integrity. The trade-off is a little extra latency and cost per gigabyte.

An MRDIMM takes a different angle. It multiplexes two ranks behind a buffer so the host interface effectively sees data at double the underlying DRAM rate. First-generation MRDIMM runs at 8800 MT/s (PC5-70400), and the JEDEC roadmap scales to 12,800 and then 17,600 MT/s toward 2030. It is aimed at bandwidth-bound workloads on Intel Xeon 6 “Granite Rapids.”

Ranks, capacities, and speeds

Rank notation — 1Rx8, 2Rx4, 4Rx4, 8Rx4 — tells you how the DRAM is organized. More ranks and x4 organization generally mean higher capacity but a heavier electrical load, which is exactly the load LRDIMM buffering exists to manage. DDR5 server speeds step from 4800 through 5600 and 6400 MT/s for RDIMM, then jump to 8800 for MRDIMM. Faster grades and denser parts are the first to go to allocation in a tight market.

Comparison table (as of 2026-08)

Prices are indicative single-unit references captured mid-2026 and move weekly; treat as reference only.

When to use each

Spec RDIMM unless you have a specific reason not to. For a balanced inference node or a general server, 64GB or 96GB DDR5-5600 RDIMMs hit the sweet spot of capacity, speed, and availability. Step to DDR5-6400 RDIMM where the platform supports it and per-core bandwidth matters.

Move to 3DS LRDIMM when you must maximize memory per socket — large in-memory datasets or fat inference nodes where 128GB modules let you fill slots without exceeding rank limits. Confirm the CPU supports the density before committing.

Choose MRDIMM 8800 only on qualified Xeon 6 platforms, and only when profiling shows memory bandwidth is the ceiling. It is a premium, allocation-controlled part, so validate both the platform HCL and real availability before designing it in. Whichever type you pick, our server memory desk can confirm live stock and second-source options.

The 2026 shortage context

DDR5 pricing is in a historic up-cycle. Contract prices rose roughly 90–95% quarter-on-quarter in Q1 2026 and 58–63% in Q2, with Q3 forecast up another 13–18% — a deceleration, not a decline (as of 2026-08). A 64GB DDR5-5600 kit is up about 485% year-on-year, and DDR5 runs roughly $12–14/GB versus about $3/GB a year earlier. Module lead times sit at 30–40+ weeks, new fab capacity is not online before late 2027, and several makers guide undersupply into 2027–2028.

Two buying implications follow. First, high-density (128–256GB), high-speed (6400), and all MRDIMM parts are effectively allocation, RFQ, or long-term-agreement supply — confirm quoted prices as “on request.” Second, verify condition and part number before ordering: OEM-branded modules cost 30–50% more but clear server warranty checks, while some channel listings mix new and used pulls. For anything beyond a handful of modules, an early RFQ beats a spot buy.

FAQ

What is the difference between RDIMM and LRDIMM? An RDIMM buffers only command and address signals through a register, keeping latency low. A 3DS LRDIMM adds a data buffer and stacks DRAM dies, which cuts the electrical load the CPU sees and lets you reach higher capacities per module — at a small latency and cost penalty.

What is MRDIMM 8800? MRDIMM (Multiplexed-Rank DIMM) multiplexes two ranks so the host sees data at roughly double the base rate. First-generation parts run at 8800 MT/s and target Intel Xeon 6 Granite Rapids, delivering about 37.5% more bandwidth than DDR5-6400 RDIMM.

Do RDIMM, LRDIMM and MRDIMM work in the same slot? No. They are physically compatible DDR5 modules but electrically distinct, and platforms qualify one type. Check the CPU and board vendor list before mixing — MRDIMM support in particular is limited to specific Xeon 6 SKUs as of 2026-08.

How high can DDR5 server memory capacity go? Single RDIMMs reach 96GB and octal-rank parts hit 256GB; 3DS LRDIMM covers the 128GB high-density tier. Micron sampled 256GB modules at up to 9,200 MT/s in May 2026, so density and speed are still climbing.

Bottom line

RDIMM is the default, 3DS LRDIMM buys density, and MRDIMM 8800 buys bandwidth on Xeon 6 — but 2026 supply, not the spec sheet, decides what you can actually get. Browse server memory options, then request a quote for live, indicative pricing and second-source guidance.