Skip to content

Dear customers! Please send orders and inquiries to [email protected] or to Telegram @servix_uz instead of calling by phone: due to the high workload, the operator cannot answer all incoming calls.

Servix

Server RAM: RDIMM, channels, ranks, DDR4 and DDR5

Servix technical team · Updated

In brief. Memory is selected not by capacity but by population: fourth- and fifth-generation Xeon Scalable processors have eight channels per processor, Xeon 6 6900P series and AMD EPYC 9004/9005 have twelve, and entry-level Xeon E has two. If not all channels are populated, part of the bandwidth is lost, and no amount of capacity can compensate for it. A second module per channel lowers the speed: on Xeon 6 it is 6400 MT/s with one module versus 5200 with two. DDR4 cannot be moved from Gen10 to Gen11, and for virtualization take headroom of at least 30% plus a reserve for node failure.

Server memory is not the same as memory for a desktop computer, and mistakes in selecting it cost either money or performance. The second mistake is more expensive: overpayment is visible immediately, while a bandwidth shortfall lives in the system for years and gets blamed on “heavy 1C” or “bad virtualization.” Here is what you need to know before ordering modules.

Module types: UDIMM, RDIMM, LRDIMM

  • UDIMM — unbuffered memory without a register. Found in entry-level servers based on Xeon E and in workstations; limited in capacity and in the number of modules per channel.
  • RDIMM — registered. The main type for servers: buffering of address signals allows more modules per channel and keeps stability at high speeds.
  • LRDIMM — with a load-reducing buffer on the data lines. Used where maximum capacity per socket is needed: the buffer takes the load off the controller and allows high-capacity modules to be used.
  • MRDIMM — a new type for Xeon 6 platforms: two ranks operate alternately, and speed rises to 8800 MT/s on the 6900P series and to 8000 MT/s on the 6700P, but only with one module per channel. It makes sense where the workload is bound specifically by memory — analytics, inference, compute workloads.

Types must not be mixed. If a server has RDIMMs, adding UDIMMs is pointless — the machine will not start. Within one type, mixing modules of different capacities and ranks is sometimes technically possible, but the platform then lowers the speed to that of the worst installed module, and some vendors simply do not support such configurations.

The main rule: populate the channels

This is where performance is most often lost. The memory controller of a server processor has several independent channels, and the total bandwidth is the sum of them. If not all channels are populated, part of the bandwidth is physically unused, no matter how many gigabytes are installed in the server.

Platform

Channels per processor

Type and speed

Sensible population

Xeon E-2300 / E-2400 (ML30, DL20, R250, T150)

2

DDR4-3200 or DDR5-4800, UDIMM ECC

2 or 4 modules, always in pairs

Xeon Scalable 2nd and 3rd Gen (Gen10, Gen10 Plus, 14G, 15G)

6 or 8

DDR4-2933 / DDR4-3200 RDIMM

6 or 8 modules per processor

Xeon Scalable 4th Gen, Sapphire Rapids (Gen11, 16G)

8

DDR5-4800 RDIMM

8 modules per processor

Xeon Scalable 5th Gen, Emerald Rapids

8

DDR5-5600 RDIMM

8 modules per processor

Xeon 6 6700P series (Gen12)

8

DDR5 up to 6400 MT/s, MRDIMM up to 8000

8 modules; second row only for capacity

Xeon 6 6900P series

12

DDR5 up to 6400 MT/s, MRDIMM up to 8800

12 modules per processor

AMD EPYC 9004 and 9005 (DL325/DL345 Gen11–Gen12, R6615, R7615)

12

DDR5 RDIMM

12 modules per processor

A practical example. A dual-socket server based on 4th Gen Xeon Scalable with 256 GB of memory can be built in two ways: 4 modules of 64 GB or 16 modules of 16 GB. The capacity is the same, but in the second case all eight channels on both processors are populated, while in the first only half are. The difference on memory-sensitive workloads — virtualization, databases, analytics — is visible to the naked eye and cannot be fixed by anything except adding modules.

Hence the consequence for expansion planning: it is better to populate all channels with smaller modules right awaythan to install two large ones and “add more later.” Adding later often requires replacing the entire set, because mixing modules of different capacities and ranks in one channel is either impossible or leads to a lower speed.

One module per channel or two: what you pay for capacity

A server usually has twice as many slots as channels: two modules per channel (2DPC) allow capacity to be doubled. The price is speed. On Xeon 6 platforms, with one module per channel memory runs at 6400 MT/s, and with two at up to 5200 MT/s; on previous generations the picture is the same, only the numbers are lower. MRDIMMs do not work in 2DPC mode at all.

Hence a simple rule: if the task is bound by memory bandwidth — analytics, inference, in-memory DBMS — take one row of higher-capacity modules. If the task is bound by capacity — virtualization with a large number of machines — a second row is justified; losing 15–⁠20% of speed costs less than a memory shortage and swapping.

Ranks and why speed drops

Modules come as single-rank (1Rx8), dual-rank (2Rx8, 2Rx4) and, for LRDIMM, quad-rank. It is worth paying attention to this for two reasons.

  • The number of ranks per channel is limited by the controller. This determines how many modules, and of what type, you can physically install — the table is in HPE QuickSpecs and in the Dell memory matrix for the specific platform.
  • Speed depends on population. When all slots are populated and multi-rank modules are used, memory may run slower than the processor's rated speed. This must be checked before ordering: the difference between 5600 and 4400 MT/s across eight channels amounts to gigabytes per second.

All else being equal, dual-rank modules provide slightly higher effective bandwidth than single-rank ones thanks to bank interleaving — when choosing between 1Rx8 and 2Rx8 of the same capacity and price, take the latter.

DDR4 or DDR5: there is no choice, only consequences

The memory type is determined by the platform, and this is one of the main reasons why a pool of modules does not move between generations.

Server generation

Memory

HPE ProLiant Gen10 / Gen10 Plus, Dell PowerEdge 14G / 15G

DDR4

HPE ProLiant Gen11 and Gen12, Dell PowerEdge 16G and newer

DDR5

Memory from Gen10 will not fit in Gen11 — different connectors, voltage and power management scheme (in DDR5 the regulator has moved onto the module itself). This is important to consider when planning a migration: a pool of DDR4 modules cannot be moved to the new platform and remains a resource only for older machines. If you plan to replace your fleet, buy only as much DDR4 as is needed to last until the migration — details in the breakdown of upgrading from Gen10 to Gen11 and Gen12.

MRDIMM: what it is and when it is needed

MRDIMM (Multiplexed Rank DIMM) — a JEDEC standard for DDR5 memory in which two ranks of a module operate simultaneously, and a data buffer (MDB) multiplexes their exchange with the controller. The effective bus speed is twice that of a single rank: the first generation delivers 8800 MT/s on the Xeon 6 6900P series and 8000 MT/s on the 6700P series, while regular DDR5 RDIMMs on the same platforms are limited to 6400 MT/s. The roadmap is 12,800 and 17,600 MT/s in the next generations. Its predecessor is the MCR DIMM from Intel and SK hynix; MRDIMM is its standardized version.

What matters in practice:

  • Platform support is required. An MRDIMM does not go into just any DDR5 slot: currently it is supported by servers based on Intel Xeon 6 with P-cores. AMD EPYC 9005 works with regular DDR5-6400 RDIMMs. In an unsupported server, the module simply does not initialize.
  • Only one module per channel. MRDIMM does not work in 2DPC mode; capacity is built up through module size: 32, 64, 96, 128 and 256 GB; the 256 GB module is produced in the tall TFF form factor for 2U chassis.
  • Do not mix with RDIMM. All modules in one system must be of the same type.
  • Price. The premium over an RDIMM of the same capacity is noticeable and justified where memory bandwidth is the bottleneck: HPC, AI inference on the CPU, in-memory DBMS, analytics. For virtualization, 1C and file servers there is no gain: there, memory is limited by capacity, not speed.

Our catalog lists DDR4 and DDR5 RDIMMs; we supply MRDIMMs made to order together with a Xeon 6 server, where such memory makes sense.

Original or compatible

Modules labeled by the server manufacturer pass its validation and are displayed correctly in the management system (iLO, iDRAC): model, serial number, corrected error counters, failure prediction. Compatible modules from specialized memory vendors are cheaper and in most cases work normally, but:

  • the management system may not show full information about the module, and sometimes may not see it in the inventory at all;
  • in a disputed case, the server manufacturer will point to the non-original component when handling a warranty claim;
  • on new DDR5 platforms with strict compatibility tables, the risk of a speed reduction is higher than it was with DDR4.

A sensible approach: original modules in critical production, compatible ones in test and auxiliary systems. The same principle applies to drives: saving on a component monitored by vendor firmware always cuts both ways.

How much memory to take for the task

  • Virtualization host: memory almost always runs out before the processor does. Take headroom of at least 30% over the calculation and account for a reserve for node failure in the cluster: in a three-node cluster, each node must be able to absorb the machines from one neighbor. A practical guideline is 512 GB per dual-socket node for 30–⁠40 machines with average load.
  • Database server: go by the size of the active part of the database — it must fit entirely in memory. A 200 GB database with a 60 GB active part runs on 128 GB of memory; the same database with full scans will require more.
  • 1C: a memory shortage shows up as “it slows down in the evening during period closing.” For 30–⁠50 users, 64–⁠128 GB is allocated, and with a dedicated DBMS server, memory for it is planned separately. A calculation by number of users is in the article server for 1C.
  • Terminal server: 2–⁠4 GB per active session plus memory for the OS itself; this is where people most often underestimate.
  • File server and backup: requirements are modest; 32–⁠64 GB is enough. The main resource here is the disk subsystem.

One more calculation point: hypervisor and OS licenses are counted by cores, not memory, so adding memory is the cheapest way to increase virtual machine density per node. A comparison of license costs is in the article Proxmox or VMware in 2026.

What to check before ordering

  • How many channels the selected processor has and how many slots the platform physically has — population is planned from channels, not from the total gigabytes.
  • The speed table for 1DPC and 2DPC for the specific server model: for HPE it is in QuickSpecs, for Dell in the memory installation guide.
  • Whether types and ranks are being mixed — when adding memory to an old server, this is the number one cause of boot failure.
  • Whether there are free slots at all: single-socket 1U servers often have 8, and half are occupied.
  • Whether there will be a second processor. Installing it requires its own set of modules — half of the channels belong to it.

Ready-made server configurations with memory already selected are in the selections HPE ProLiant Gen11 and Dell PowerEdge. What makes up the price of a server in general is covered here.

Questions and answers

Can desktop computer memory be installed in a server?

No. Server platforms require error-correcting memory (ECC), and higher-end ones require registered memory (RDIMM), which is electrically incompatible with regular modules. Even where a non-ECC UDIMM physically fits, the platform will either not accept it or will run without protection against single-bit errors — and that is exactly the class of failures that makes a server reboot once a month for no clear reason.

What is more important — capacity or speed?

For virtualization and terminal servers, capacity is more important: a memory shortage causes swapping, which slows the system down several-fold. For analytics, inference and DBMS with heavy scanning, bandwidth matters more, meaning populated channels and high speed. With an equal budget, populate all channels first, then increase capacity.

Can memory be added to a server that is already in operation?

Yes, if the type, rank, speed and capacity of the modules match and there are free slots in the correct positions. The slot population order for HPE and Dell is strictly defined and differs from “left to right” — with incorrect installation, the server will either not see part of the memory or will lower the speed. Send us the server model and serial number, and we will check compatibility against the configuration free of charge.

Is higher-speed memory needed if the processor is entry-level?

No: the processor's controller caps the speed, and faster modules will simply run at the platform's rated speed. Paying extra for MT/s makes sense only when both the processor and the population (one module per channel) support it.

We will select modules compatible with your specific platform, checked against the server configuration and slot population order — write to us, [email protected] or Telegram.

Illustration: smial, Wikimedia Commons (Free Art License).

Related reading