Server Memory Buyer's Guide

Rank and Density — The First Compatibility Filter

DIMM rank refers to the number of independent memory banks on a module. Single-rank (1R) DIMMs have one bank; dual-rank (2R) have two; quad-rank (4R) have four. Higher rank gives more banks the memory controller can interleave, improving throughput. But higher rank also reduces the number of DIMMs per channel (DPC) the controller can drive at full speed.

For most modern Intel Xeon and AMD EPYC platforms: 1R and 2R DIMMs can populate 1-2 DIMMs per channel at full rated speed. 4R DIMMs typically downclock at 2 DPC. LRDIMMs (load-reduced DIMMs) decouple the rank constraint by using a memory buffer on the module itself, allowing 8R logical configurations to run at full speed.

Density is the per-module capacity (8GB, 16GB, 32GB, 64GB, 128GB). Higher density modules cost more per GB but require fewer slots. Higher density also means higher rank (typical correlation: 8GB = 1R, 16-32GB = 2R, 64-128GB = 2R or 4R).

Rule of thumb: match rank and density across all DIMMs in the same channel. Mixed configurations force the memory controller to fall back to the lowest-common-denominator timings.

Channel Population — Why Half-Populated Servers Lose 50% of Bandwidth

Modern Xeon Scalable and EPYC platforms have 6-12 memory channels per CPU socket. Memory bandwidth scales linearly with populated channels: a 12-channel platform populated with 6 DIMMs delivers half the bandwidth of the same platform populated with 12 DIMMs.

For a 2-socket server with 12 channels per CPU (24 channels total), the population sequence is: 6 DIMMs per CPU first (one per channel), then 12 DIMMs per CPU (two per channel). Asymmetric population — say 4 DIMMs in CPU 0 and 6 in CPU 1 — produces NUMA imbalance and is detected by VMware as a configuration warning.

Concrete recommendation: never half-populate the memory subsystem unless you are explicitly trading bandwidth for capacity headroom. For most workloads, fully populating with smaller DIMMs at full bandwidth outperforms half-populating with larger DIMMs.

ECC, RDIMM, LRDIMM, Unbuffered — Pick the Right Type

ECC memory adds an extra chip per rank that stores parity bits for single-bit error correction and multi-bit error detection. Server-grade memory is essentially always ECC; non-ECC server memory exists but creates support headaches and is rarely worth the cost savings.

RDIMM (Registered DIMM) is the dominant server memory type. The "Registered" buffer chip sits between the memory controller and the DRAM chips, isolating capacitive load. RDIMMs work in nearly every modern server motherboard.

LRDIMM (Load-Reduced DIMM) adds a full memory buffer on the DIMM, allowing higher-density modules at full speed. LRDIMMs typically cost 20-40% more than RDIMMs of equivalent capacity but enable configurations that pure RDIMMs cannot (e.g., 12 x 128GB LRDIMMs on a single CPU socket).

UDIMM (Unbuffered DIMM) is desktop/workstation memory. Server motherboards typically reject UDIMMs at boot. Do not buy UDIMMs for server platforms.

OEM QVL — Why Validated Memory Matters

Every server OEM publishes a Qualified Vendor List (QVL) for memory — a list of specific memory part numbers validated against each server model. The QVL is the authoritative source for "will this memory work in my server."

For Dell PowerEdge, the QVL is searchable by service tag at dell.com/support. The PERC controller behavior may also pre-validate memory in newer Gen14+ servers.

For HPE ProLiant, the SmartMemory line is the OEM-branded memory tier. SmartMemory modules carry an HP-specific identifier that iLO recognises; non-SmartMemory modules work but generate iLO warnings on some Gen10+ servers.

For Lenovo ThinkSystem, the part-number prefix (4ZC, 7X77, etc.) indicates compatibility class. The Lenovo ServerProven program lists validated configurations.

For whitebox / Supermicro, the motherboard manual carries the QVL. Supermicro QVLs are typically broader than the major OEMs because the company tests across many third-party memory brands.

Pricing Tier — New, Bulk-OEM, Refurbished

Server memory at Pro Disk Network ships across three tiers. The tier choice is driven by workload criticality and warranty preference.

New (Sealed OEM): factory-sealed retail-channel memory carrying the full OEM warranty (typically 5 years for HPE SmartMemory, lifetime for Crucial enterprise, varying terms for Samsung/Micron OEM). Premium for tier-1 production databases and HA cluster build-outs. Cost: 100% baseline.

New (Bulk OEM): factory-fresh memory sold without retail box, carrying a 36-month Pro Disk Network warranty. Cost-optimal for hyperconverged clusters, virtualisation hosts, and general-purpose servers. Cost: 60-75% of sealed OEM.

Refurbished: memory pulled from decommissioned servers, individually MemTest-validated, carrying a 12-month advance-replacement warranty. Appropriate for development, lab, dev/test, and capacity expansion on aging hardware. Cost: 30-45% of sealed OEM.

Pro Disk Network ships matched-lot refurbished memory (sequential manufacturing dates) on bulk orders of 16+ modules — useful for hyperconverged clusters where DIMM rank consistency matters.