UK’s trusted IT infrastructure partner since 2003
sales@servnetuk.com
0800 987 4111
Servnet
ConfiguratorGet in Touch
Computing Components · Memory & RAM

Server Memory — DDR5 RDIMM, LRDIMM & ECC
Complete buyer guide & specification reference

Server memory is the most common hardware decision made without enough information — wrong DIMM type, wrong speed, or under-populated channels are all silent performance killers. This guide explains every key concept, compares RDIMM vs LRDIMM, identifies what each workload actually needs, and documents compatibility for Dell, HPE, Lenovo, and Supermicro platforms.

Get Memory Quote →Match to your CPU →
DDR5
Standard on all new Intel Xeon 6 & EPYC 9005 platforms
12 ch
AMD EPYC 9005 memory channels — highest in x86
6TB+
Max per 2P server — 24× 256GB LRDIMM
5600 MT/s
DDR5 native speed on EPYC Turin & Xeon 6
ECC
Mandatory for all server platforms — not optional
Lifetime
Warranty — Samsung, Micron, Kingston server DIMMs
Understanding Server Memory

Why server RAM is fundamentally different from desktop RAM

Desktop and laptop systems use UDIMM (Unbuffered DIMM) — a simple, low-cost design with the DRAM chips connected directly to the CPU memory controller. This works fine when you have 2–4 DIMMs per system, but at server densities (8–24 DIMMs per CPU) the electrical load on the memory bus causes signal integrity problems and crashes.

Enterprise servers solve this by placing a register or buffer chip on the DIMM itself. This register isolates the DRAM chips from the memory bus, dramatically reducing the electrical load. The CPU only "sees" the register, not all 18+ DRAM chips behind it. This is what makes RDIMM (Registered DIMM) the universal standard for server memory.

DDR5 introduced on-DIMM power management (PMIC), doubled the internal burst length to 16, and moved to a differential clock architecture — all of which improve signal integrity at higher speeds. This is why DDR5 can scale to 6400 MT/s+ natively whereas DDR4 topped out at 3200 MT/s in server use.

DDR4 vs DDR5: Key Differences for Servers

Voltage1.2V1.1V
Max speed (server)3200 MT/s6400 MT/s+
Channel width64-bit2× 32-bit sub-channels
On-DIMM power mgmtNoYes (PMIC)
ECC granularityDIMM-levelPer sub-channel
Max DIMM capacity128GB RDIMM256GB+ LRDIMM
SpecificationDDR4DDR5 ↑

Key rule: DDR5 is the only option for Intel Xeon 6 and AMD EPYC 9005 platforms. DDR4 is only relevant for older Intel Xeon 4th Gen (Sapphire Rapids) and AMD EPYC 9004 (Genoa) systems that were shipped with DDR4 configuration options.

DDR5 server memory selector — RDIMM, 3DS RDIMM, MRDIMM, LRDIMM compared by capacity, speed, and workload fit
Key Terminology

Server Memory Concepts Explained

These are the terms that determine what memory your server can actually use — getting them wrong results in incompatibility or silent performance loss.

🛡️
ECC
Error-Correcting Code

ECC memory detects and corrects single-bit memory errors in real time, preventing data corruption and system crashes. All enterprise server memory uses ECC — it is not optional. A server running without ECC will silently produce incorrect results when bit-flip events occur (typically caused by cosmic rays or electrical noise).

🔧
RDIMM
Registered DIMM

A register (also called a "buffer") on the DIMM sits between the CPU memory controller and the DRAM chips, reducing the electrical load on the memory bus. This allows more DIMMs per channel and higher capacities. All mainstream enterprise servers use RDIMMs. They operate at standard DDR5 speeds.

📦
LRDIMM
Load-Reduced DIMM

Goes further than RDIMM by also buffering the data signals (not just the address/command signals). This allows quad-rank and octa-rank DIMMs to operate without overloading the memory controller, enabling 256GB per slot. The trade-off is slightly higher latency and a small bandwidth reduction vs RDIMM at the same speed.

DDR5 Speed
Data Rate (MT/s)

DDR5-4800 means 4,800 megatransfers per second. DDR5-5600 means 5,600 MT/s — 17% faster. The CPU platform determines the maximum speed: AMD EPYC 9005 (Turin) runs DDR5-5600 natively; Intel Xeon 6 E-core runs DDR5-5600; Intel Xeon 6 P-core supports DDR5-6400 with HBM options. Running faster DIMMs in an older platform just drops them to the platform max.

🔀
Memory Channels
Independent Bus Paths

Each memory channel is an independent 64-bit bus between the CPU and DRAM. More channels = more parallel bandwidth. AMD EPYC 9005 has 12 channels (leading x86); Intel Xeon 6 P-core has 12 channels; E-core has 8. Populating all channels symmetrically is critical — an unbalanced configuration reduces bandwidth to that of the lowest-populated channel.

📋
Rank
1R / 2R / 4R / 8R

A "rank" is a set of DRAM chips that can be addressed simultaneously. Single-rank (1R) DIMMs are faster to access but lower capacity. Dual-rank (2R) adds capacity. Quad-rank and octa-rank (4R/8R) are only found in LRDIMMs. More ranks per channel can reduce maximum achievable speed — check your server's Memory Population Guide for limits.

DIMM Type Comparison

UDIMM vs RDIMM vs LRDIMM — Which Do You Need?

Server platforms specify exactly which DIMM type they support. Using the wrong type will prevent the server from booting. RDIMM is the correct choice for 99% of enterprise servers — LRDIMM only when you genuinely need 256GB per slot.

TypeBufferMax CapacityRanksECCLatencyBest For
UDIMMNone32GB1R/2ROptionalLowestWorkstations only — not for servers
RDIMMAddress/CMD buffer128GB1R/2R/4RRequiredLowAll enterprise servers — universal choice
LRDIMMFull data + addr buffer256GB4R/8RRequiredSlightly higherSAP HANA, AI serving, max density builds
Workload Selection Guide

How Much Memory Does Your Workload Actually Need?

Over-specifying wastes budget; under-specifying throttles performance or causes OOM crashes. This guide gives a starting point — always verify against your actual dataset and VM count before finalising.

WorkloadRecommended DIMMReasoningSpeed Target
VMware ESXi / Hyper-V Virtualisation32–64GB RDIMMAllocate 4–8GB per VM plus 16GB for hypervisor overhead. A 2P server with 24× 64GB RDIMM provides 1.5TB — comfortable for 100–200 VMs.DDR5-4800
In-Memory Database (Redis, SAP HANA)128–256GB RDIMM/LRDIMMDatabase must fit entirely in DRAM. SAP HANA requires a specific memory-to-CPU ratio. 256GB LRDIMMs maximise capacity in 24-slot servers to 6TB.DDR5-4800 (LRDIMM)
AI/ML Inference Server64–128GB RDIMMModel weights must load from RAM to GPU. 128GB × 8 channels = 1TB/s+ aggregate bandwidth improves data feeding for AI workloads. DDR5-5600 preferred.DDR5-5600
High-Performance Computing (HPC)64–128GB RDIMMNUMA-aware applications benefit from equal DIMM population across all channels. Prioritise bandwidth over capacity — fewer, faster DIMMs outperform more, slower ones for HPC.DDR5-5600
AI Model Serving (LLMs, 70B+ params)256GB LRDIMMLarge language models require system RAM for CPU-side tokenisation, context management, and KV cache. 6TB in a 2P server supports concurrent serving of multiple large models.DDR5-4800 (LRDIMM)
Web / App / Containerised Workloads32GB RDIMMContainer workloads typically allocate 1–4GB per container. Single-rank 32GB DIMMs are cost-optimal and achieve rated speeds more easily than dual-rank.DDR5-4800
Platform Compatibility

Memory Specifications by CPU Platform

The CPU determines maximum DDR5 speed, channel count, and DIMM capacity. Running faster DIMMs than the platform supports simply drops them to the platform maximum — you pay for speed you cannot use.

CPU PlatformChannelsMax DDR5 SpeedMax Per DIMMMax Per 2PNotes
AMD EPYC 9005 (Turin)12DDR5-5600256GB LRDIMM6TBHighest channel count — populate all 12 channels for max bandwidth
Intel Xeon 6 P-core12DDR5-6400 (w/ HBM option)256GB LRDIMM6TBBirch Stream platform supports optional HBM alongside DDR5
Intel Xeon 6 E-core8DDR5-5600128GB RDIMM3TBP4S platform; 8 channels vs 12 means 33% less bandwidth than P-core
Intel Xeon 5th Gen (Emerald Rapids)8DDR5-5600256GB LRDIMM6TBEagle Stream platform — supports DDR5 and DDR4 in specific configs
AMD EPYC 9004 (Genoa)12DDR5-4800256GB LRDIMM6TBPrevious gen — max DDR5-4800; Turin upgrade path keeps same SP5 socket
Available Modules

DDR5 Server Memory Specifications

All specifications are manufacturer-published. Part numbers are OEM-independent — QVL-verified for Dell PowerEdge, HPE ProLiant, Lenovo ThinkSystem, and Supermicro platforms.

🧠
Samsung · RDIMM
M321R8GA0BB0-CQK

Samsung 64GB DDR5-4800 RDIMM

Capacity64 GB
SpeedDDR5-4800 (PC5-38400)
Rank2Rx4
Voltage1.1V
ECCYes
CAS LatencyCL40
CompatibleIntel Xeon 4th/5th Gen · AMD EPYC 9004/9005
Best for: Enterprise virtualisation, database servers, general-purpose workloads.
🧠
Samsung · RDIMM
M321R8GA0PB0-CWM

Samsung 64GB DDR5-5600 RDIMM

Capacity64 GB
SpeedDDR5-5600 (PC5-44800)
Rank2Rx4
Voltage1.1V
ECCYes
CAS LatencyCL46
CompatibleIntel Xeon 6 · AMD EPYC 9005 (Turin) — DDR5-5600 native
Best for: AMD EPYC Turin and Intel Xeon 6 at native DDR5-5600 — AI inference, analytics.
🧠
Micron · RDIMM
MTC40F2046S1RC56BA1R

Micron 128GB DDR5-5600 RDIMM

Capacity128 GB
SpeedDDR5-5600 (PC5-44800)
Rank2Rx4
Voltage1.1V
ECCYes
CAS LatencyCL46
CompatibleIntel Xeon 6 · AMD EPYC 9005 · Dell R760 · HPE DL380 Gen11
Best for: High-density virtualisation, in-memory databases — 3TB+ per 2P server (24 DIMMs × 128GB).
🧠
Samsung · LRDIMM
M321RAGA0B20-CWK

Samsung 256GB DDR5-4800 LRDIMM

Capacity256 GB
SpeedDDR5-4800 (PC5-38400)
Rank8Rx4 via LRB buffer
Voltage1.1V
ECCYes
BufferLoad Reduce Buffer (LRB)
CompatibleIntel Xeon 6 P-core · AMD EPYC 9005 (24-DIMM platforms)
Best for: Maximum per-server capacity — 6TB in 24-DIMM 2P system. SAP HANA, LLM serving, in-memory analytics.
🧠
Micron · RDIMM
MTC40F204WS1RC48BA1R

Micron 32GB DDR5-4800 RDIMM

Capacity32 GB
SpeedDDR5-4800 (PC5-38400)
Rank1Rx4
Voltage1.1V
ECCYes
CAS LatencyCL40
CompatibleIntel Xeon 4th Gen · AMD EPYC 9004 · Dell R660/R760 · HPE DL360/DL380
Best for: Mainstream servers — cost-effective entry for virtualisation, web, and application workloads.
🧠
Kingston · RDIMM
KSM56R46BD4PMI-96HMI

Kingston 96GB DDR5-5600 RDIMM

Capacity96 GB
SpeedDDR5-5600 (PC5-44800)
Rank2Rx4
Voltage1.1V
ECCYes
CAS LatencyCL46
CompatibleAMD EPYC 9005 Turin · Intel Xeon 6 — native DDR5-5600 platforms
Best for: Mid-density configs — 96GB fills the gap between 64GB and 128GB per slot.
Common Questions

Frequently Asked Questions

Q: Does ECC memory slow down a server?
In practice, no. The ECC overhead is negligible — benchmarks show less than 1% performance difference versus non-ECC at the same speed grade. The stability and data integrity benefit far outweighs this imperceptible overhead for any production workload.
Q: Can I mix DDR5-4800 and DDR5-5600 DIMMs in the same server?
Technically possible but not recommended. When mixing speeds, the memory controller drops all DIMMs to the slowest speed — you lose the bandwidth advantage of the faster DIMMs entirely. For best results, populate all slots with the same part number.
Q: What happens if I populate memory asymmetrically?
Asymmetric population (different DIMMs per channel, or some channels empty) reduces bandwidth to the lowest-populated channel count. On an AMD EPYC 9005 with 12 channels, populating only 8 channels leaves 33% of the memory bandwidth unused. Always follow the server's Memory Population Guide.
Q: RDIMM vs LRDIMM: should I always choose LRDIMM for maximum capacity?
Not necessarily. LRDIMMs are slightly slower (the LRB buffer adds ~2–3ns latency) and cost more. For workloads where 128GB RDIMM meets your capacity need, RDIMM will be faster and cheaper. Only move to LRDIMM when you genuinely need 256GB per slot — typically for SAP HANA, large AI model serving, or databases that cannot otherwise fit in memory.
Q: Do I need to buy "branded server memory" from Dell/HPE/Lenovo, or can I use OEM-independent DIMMs?
Major server OEMs (Dell, HPE) maintain Qualified Vendor Lists (QVLs) and their servers will check memory compatibility at POST. Samsung, Micron, and Kingston manufacture server DIMMs that are QVL-listed for all major platforms — these are functionally identical to the branded variants sold by the server OEMs, often at lower cost.
Q: How much memory does an AI inference server need?
The model weights must fit in VRAM (GPU memory), but the host system RAM handles tokenisation, batching, the KV cache, and OS overhead. A practical rule: allocate at least 1.5× the GPU VRAM in system RAM. For a server with 8× H100 80GB (640GB VRAM total), 1TB of system RAM (8× 128GB DDR5) is a reasonable baseline.

Related Products & Platforms

Need memory for a specific server model?

Provide your server model and current configuration — we confirm the compatible part number, available capacity options, and UK lead time.

Request Memory Quote