Server Memory Types, Specs & Speeds Explained
Understand the specifications that identify server memory: DDR generation, RDIMM, LRDIMM or MRDIMM type, capacity, speed, ECC, rank, and chip width. These specs let you compare modules, while exact compatibility still depends on the server platform, population rules, OEM qualification, and manufacturer part number.
Jump to a memory specification
Spec cheat sheet
| Spec | Example | Why it matters |
|---|---|---|
| Generation | DDR4 | Platform generation — sets voltage, speed range, socket compatibility |
| Module type | RDIMM | Module architecture — buffering, capacity ceiling, latency |
| Capacity | 32GB | Capacity per DIMM |
| Speed | 3200 MT/s | Rated transfer rate — actual speed depends on the platform |
| Rank / width | 2Rx8 | Organization — affects exact compatibility, not just market class |
| ECC | ECC | Error correction — required for production server workloads |
| Part number | M393… | Exact product identity — the only field that guarantees an exact match |
Put together, these read as a single spec string, e.g. 32GB DDR4-3200 ECC 2Rx8 RDIMM.
The core specs that define a market class
- Generation — DDR3 / DDR4 / DDR5 — determines voltage, speed range, and socket compatibility
- Form factor — RDIMM / LRDIMM / UDIMM / MRDIMM — the buffer type, which determines capacity ceiling and latency profile
- Capacity — 16 GB / 32 GB / 64 GB / 128 GB — total usable capacity per module
- Speed — Expressed in MT/s (megatransfers per second). CPU memory controller determines the ceiling
- ECC support — Yes / No — error-correcting code. Required for any production server workload
- Voltage — 1.5 V (DDR3) / 1.35 V (DDR3L) / 1.2 V (DDR4) / 1.1 V (DDR5)
Market class answers: “What kind of module is this economically comparable to?”
Exact SKU/platform compatibility answers: “Can I safely use this exact module in this server?”
These specs are enough to place a module in a market class for comparison and pricing. They are not enough on their own to guarantee exact compatibility — that also depends on rank, x4/x8 device organization, server/platform support, CPU generation, DIMMs per channel, memory population rules, OEM qualification, and the exact manufacturer part number. Always confirm against your server vendor’s memory compatibility documentation before buying for a specific system.
Generation
| Generation | Years active | Voltage | Low-voltage variant | Typical speed range | Typical capacity range |
|---|---|---|---|---|---|
| DDR3 | 2007–2018 | 1.5 V | DDR3L — 1.35 V | 1066–1866 MT/s | 4–32 GB per module |
| DDR4 | 2014–ongoing | 1.2 V | — | 2133–3200 MT/s | 8–128 GB per module |
| DDR5 | 2021–ongoing | 1.1 V | — | 4800–8800 MT/s | 16–256 GB per module |
DDR3 is end-of-life but still active in the secondary market. DDR4 is the current procurement mainstream. DDR5 is required for Intel Sapphire Rapids+ and AMD Genoa+ platforms.
Server memory types: RDIMM, LRDIMM, UDIMM & MRDIMM
RDIMM — Registered DIMM
A register buffer sits between the memory controller and the DRAM chips, reducing the electrical load on the controller. Standard for dual-socket and higher-density server configurations. The buffer adds 1–2 clock cycles of latency — not measurable in production workloads. RDIMMs are the default assumption for any enterprise server procurement.
LRDIMM — Load-Reduced DIMM
Adds a full isolation buffer (not just a register), which reduces the electrical load further and enables more DRAM chips per module. This is historically how the highest-capacity single DIMMs (128 GB+) have been built, though DDR5 3DS-stacked RDIMMs have narrowed that gap at some capacities — check the module’s labeled form factor rather than assuming it from capacity alone. The trade-off is higher latency than RDIMM. Use LRDIMM when you need maximum memory density per server; use RDIMM when latency matters more than raw capacity.
UDIMM — Unbuffered DIMM
No buffer. The memory controller connects directly to the DRAM chips. Lowest latency and lowest cost, but a lower capacity ceiling and reduced signal integrity at higher speeds. Found in single-socket workstations, edge appliances, and budget servers. Generally not supported in dual-socket server platforms — verify against the specific system’s memory support documentation rather than assuming.
MRDIMM — Multiplexed Rank DIMM
A DDR5-era standard that multiplexes two ranks to effectively double bandwidth per channel. Intel Xeon 6 (Granite Rapids) was the first platform to support MRDIMMs natively. Adoption is still concentrated on the newest platform generations — confirm both module and platform support before specifying it, and don’t assume it’s a drop-in replacement for RDIMM.
SODIMM — Small Outline DIMM
Physically smaller than a standard DIMM — used in laptops and compact embedded systems. Appears in servers only in edge appliances, micro-servers, and some 1U platforms. Not a standard enterprise server form factor.
Server memory speed and MT/s bins
Memory speed is expressed in MT/s (megatransfers per second). More MT/s = more bandwidth per channel, which matters for memory-bandwidth-bound workloads (in-memory databases, HPC, AI inference). The CPU memory controller sets the ceiling — a faster module installed in a slower controller runs at the controller’s max, not the module’s rated speed. Always match module speed to the platform’s supported speed bins to avoid leaving bandwidth on the table.
| Generation | Common speed bins (MT/s) |
|---|---|
| DDR3 | 1066 · 1333 · 1600 · 1866 |
| DDR4 | 2133 · 2400 · 2666 · 2933 · 3200 |
| DDR5 | 4800 · 5200 · 5600 · 6000 · 6400 · 6800 · 7200 · 8800 |
DDR4-3200 and DDR5-4800 are the typical base speeds for current-generation platforms. Higher DDR5 speeds (6400+) require platform validation — check your CPU’s QVL.
What server memory speed do I need?
- Every DIMM has a rated maximum speed — e.g. a DDR4-3200 module is rated to run at 3200 MT/s.
- The server/CPU platform's memory controller determines the speed actually achievable, not the module's rating.
- DIMMs-per-channel (1DPC vs. 2DPC) and how the slots are populated can reduce achievable speed below what any individual module supports.
- A faster-rated module installed on a platform with a lower ceiling will run at the platform's supported speed, not its own rated maximum.
- In mixed configurations, the system typically runs according to the supported configuration as a whole rather than each module's individual maximum — exact behavior varies by platform and vendor.
Before buying for a specific server, check that server manufacturer’s memory population/configuration documentation (QVL) for the exact speed your configuration will actually run at.
Rank, chip width, and die density
These specs are mostly abstracted away from buyers, but they surface in compatibility sheets and occasionally affect pricing.
| Spec | What it means | When it matters |
|---|---|---|
| Rank (1R / 2R / 4R) | How many independent groups of chips the memory controller addresses on one module. Single-rank (1R) has lower load; dual-rank (2R) enables interleaving for higher throughput. | High-density configs: some platforms limit total ranks per channel, so a 4-rank LRDIMM may not be compatible where a 2-rank module is. |
| Chip width (×4 / ×8 / ×16) | How many bits each DRAM chip contributes per access. ×4 chips need more chips to fill a 64-bit bus but give finer ECC granularity. | ×4 chips are required for Advanced ECC (Chipkill/SDDC) on some server platforms. Check the platform QVL if running demanding RAS configs. |
| Die density | The capacity of each individual DRAM chip (e.g. 8Gbit, 16Gbit). Higher die density enables higher module capacity with the same chip count. | Mostly invisible to buyers. Affects whether a new-generation high-density module is supported on older platforms with density limits. |
ECC server memory and why it matters
ECC (Error-Correcting Code) memory detects and corrects single-bit memory errors in hardware, transparently to the running application — no crash, no data corruption. That doesn’t mean the event is invisible at the system level: most enterprise platforms still log corrected-error counts via IPMI/BMC or OS-level tooling (e.g. EDAC, mcelog) so operators can monitor for a failing DIMM. ECC also detects (but cannot correct) multi-bit errors and can flag them before they corrupt data. Required for any production server workload. Non-ECC memory is not a real trade-off in enterprise contexts — the cost difference is negligible and the risk is not. If a spec sheet doesn’t explicitly say ECC, assume it doesn’t have it. All RDIMM and LRDIMM modules are ECC by definition.
Market class vs. exact part number
MarketSignalIndex tracks two related but different questions about every module.
Market class
The specs above are enough to group economically comparable modules for pricing:
- DDR generation
- Form factor / module type
- Capacity
- Rated speed
Exact SKU / MPN
Identifying the exact product — and evaluating whether it fits a specific server — takes more:
- Manufacturer
- Rank / device organization
- OEM qualification
- Platform support
- Other module-specific attributes
Not every module inside a market class is automatically interchangeable — two DDR4-3200 32 GB RDIMMs in the same market class can still differ in rank, qualified platforms, or OEM lock-in. That’s why MarketSignalIndex has both market-class pages — grouped by generation, form factor, capacity, and speed — and exact part-number pages for the individual SKUs trading within them. Market class tells you what to compare on price; the exact part number tells you what you’re actually buying.
Reading a server memory part number
Samsung M393A2K40DB3-CWE — DDR4 RDIMM 16 GB 3200
| Segment | Value | Meaning |
|---|---|---|
| Product family | M393 | Samsung server DRAM, DDR4 RDIMM (M=Memory, 39=DDR4 server, 3=RDIMM) |
| Die generation | A | Internal die revision letter — increments with each process shrink |
| Configuration | 2K40 | Capacity and rank encoding (internal Samsung code — 16 GB 2Rx8 in this case) |
| PCB revision | DB3 | Board design iteration |
| Speed bin | CWE | C = DDR4-3200, W = 1.2 V, E = ECC / unbuffered variant code |
SK Hynix HMAA4GR7AJR8N-XN — DDR4 RDIMM 32 GB 3200
| Segment | Value | Meaning |
|---|---|---|
| Brand | H | SK Hynix |
| Product type | MA | DRAM module (MA = DDR4 server module prefix) |
| Die generation | A | Internal die revision |
| Die density | 4G | 4 Gbit per DRAM chip |
| Form factor | R | RDIMM |
| Generation | 7 | DDR4 (Hynix internal generation code) |
| Configuration | AJR8N | Rank, width, and internal config code |
| Speed bin | XN | XN = DDR4-3200 |
Micron MTC20F2085S1RC56BR — DDR5 RDIMM
| Segment | Value | Meaning |
|---|---|---|
| Brand | MT | Micron Technology |
| Product type | C | DRAM component / module type indicator |
| Generation | 20 | DDR5 (Micron uses 20 for DDR5 module series) |
| Configuration | F2085S1R | Internal density, rank, and width encoding |
| Speed bin | C56 | DDR5-5600 |
| Revision | BR | Hardware revision |
Vendor part number schemes are partially proprietary — treat the segment breakdown as directional. The confirmed specs (capacity, speed, form factor, generation) are always on the datasheet and product label; the internal codes vary by product family.
How this connects to pricing
Because Samsung, SK Hynix, and Micron all build to the same JEDEC spec, secondary market prices within the same market class are highly correlated across brands — typically within 5–10%. A Samsung DDR4-3200 32 GB RDIMM and a Hynix DDR4-3200 32 GB RDIMM will usually trade within a few dollars of each other. Brand premium in memory is real but small compared to CPUs or HDDs, where brand maps to a distinct product. The bigger price driver in memory is the market class itself: generation, capacity, and speed bin — exact platform fit is a separate question from price.
Check server memory market signals
Everything above is enough to place a module in a market class. MarketSignalIndex tracks live market signals for server memory organized the same way — browse by generation, module type, capacity, and speed, then drill down into the exact SKUs and part numbers currently trading in that market.