Server RAM is registered (RDIMM) or load-reduced (LRDIMM) memory with ECC (error-correcting code). The consumer-grade unbuffered DIMM (UDIMM) does not work in servers. The right pick is RDIMM with ECC for the vast majority of deployments; LRDIMM for the highest-capacity slots. The team that orders memory for a server checks the CPU’s memory controller spec first, then matches speed, rank, and capacity to the workload.
Table of contents
- UDIMM vs RDIMM vs LRDIMM
- ECC and why servers need it
- Rank, speed, and capacity
- How to identify what your server has
- How this fits the rest of the stack
- FAQ
UDIMM vs RDIMM vs LRDIMM
UDIMM (Unbuffered DIMM) - the memory in laptops and desktops. No buffer chip between the CPU and the DRAM chips. Cheaper, lower latency, but limited to about 2 DIMMs per channel and no ECC. The team that puts UDIMM in a server gets a POST failure or a noisy log of memory errors.
RDIMM (Registered DIMM) - the standard server memory. A register buffer chip sits between the CPU and the DRAM, which lets the memory controller drive more DIMMs per channel (4-8 typical) and supports ECC. The latency hit from the register is in the single-nanosecond range, so it is invisible in real workloads.
LRDIMM (Load-Reduced DIMM) - the high-capacity server memory. A larger buffer reduces the electrical load on the memory controller, allowing 8 DIMMs per channel at high speeds. The team that needs 1 TB+ per socket uses LRDIMM.
ECC and why servers need it
ECC (Error-Correcting Code) memory detects and corrects single-bit errors and detects multi-bit errors. A consumer machine with a flipped bit is a random crash. A server with a flipped bit - say, in a database page or a TLS handshake - is a silent corruption.
Server-class CPUs (Intel Xeon, AMD EPYC) support ECC. Consumer CPUs (Intel Core, AMD Ryzen) generally do not. The team that builds on Xeon or EPYC gets ECC. The team that builds on consumer silicon cannot - the memory controller does not support it.
ECC is not optional for production. The team that runs anything customer-facing uses ECC memory.
Rank, speed, and capacity
Rank - the number of independent 64-bit data sets on a DIMM. Single-rank (1R), dual-rank (2R), quad-rank (4R). Higher rank = more capacity per DIMM, but the memory controller can drive fewer ranks per channel at high speed. The team that has 8 slots per channel and 16 slots of capacity goes 1Rx8 or 2Rx8, not 4Rx4.
Speed - the data rate, in MT/s (mega-transfers per second). DDR4-3200, DDR5-4800 are common. The team that buys the fastest speed the CPU supports pays a premium; the team that buys the JEDEC-standard speed (e.g. DDR5-4800 for a Xeon) gets the best price/GB.
Capacity - the GB per DIMM. 16 GB, 32 GB, 64 GB, 128 GB, 256 GB. The team that needs 1 TB per server fills 8 slots with 128 GB DIMMs (or 4 slots with 256 GB LRDIMMs).
How to identify what your server has
Linux tools:
sudo dmidecode -t memory | head -50
This prints the manufacturer, speed, type, and rank of every DIMM in the system. The team that diagnoses a memory error runs this first.
lshw -short -C memory
A short summary. The team that needs a quick check runs this.
FAQ
Can I mix UDIMM and RDIMM?
No. The memory controller detects the DIMM type at POST. Mixing is either rejected outright or causes intermittent errors. The team that tries to save a few dollars by mixing types gets a server that does not boot.
Do I need ECC memory?
Yes, for any production server. ECC catches single-bit errors that would otherwise become silent corruption in databases, file systems, or TLS handshakes. Consumer machines do not need ECC; servers do.
What is the difference between DDR4 and DDR5?
DDR5 is the newer generation - higher speeds (4800 MT/s and up), higher capacity per DIMM (256 GB and beyond), and on-DIMM voltage regulation. DDR4 tops out at 3200 MT/s and 128 GB per DIMM. The team that buys a new server buys DDR5; the team that extends an existing fleet checks the CPU’s memory controller.
How much RAM do I need?
The rule of thumb: enough to hold the working set with 20-30% headroom. For a database server, the working set is the hot indexes and the hot pages - usually measured with pg_buffercache or the database’s cache inspector. For an application server, the working set is the application code plus the in-process cache.
What happens if I install a DIMM that is too slow?
The system runs the DIMM at the slower speed, not the speed the DIMM is rated for. A DDR4-2666 in a DDR4-3200 system runs at 2666. The team that buys mixed speeds gets the slowest common denominator.
How this fits the rest of the stack
For a sense of what the full project costs before it commits, the RunxBuild hosting calculator shows the line items together. The API, the database, the storage, the worker, the bandwidth - each one is a separate number, and the team’s mental model for the platform is the sum of those numbers.
Useful related references: