ECC Memory Explained: Four Different Mechanisms, One Word

Quick answer

"ECC" describes four genuinely different things depending on what carries it: plain non-ECC (no error correction at all), ECC UDIMM (error correction on an unbuffered module), registered ECC (error correction combined with an RDIMM's register buffer), and DDR5's on-die ECC (automatic per-chip correction inside the DRAM package itself, which sits alongside - not in place of - the module's own system-level ECC bits, per Lenovo Press documentation).

Updated · vendor documents read 14-16 September 2026

Share:WhatsAppFacebookPost on X

Non-ECC: no correction bits at all

A non-ECC module carries only data bits - no extra circuitry to detect or correct a bit flip. Lenovo Press's own DDR5 documentation states the comparison directly: "Non-ECC DIMMs have only the 32 data bits per channel, and will only be used in end-user devices." That last clause is doing real work: it positions non-ECC as the consumer-desktop default, not something a server platform's own memory documentation typically lists as a supported option.

ECC UDIMM: correction without the register buffer

An ECC UDIMM adds the extra correction bits to an otherwise unbuffered module - the signal path from controller to DRAM stays direct, exactly as on a plain UDIMM, with correction bits layered on top. It solves the data-integrity problem without solving the electrical-load problem an RDIMM's register buffer addresses, which is why it shows up mainly on entry-level towers and workstations rather than on high-density rack servers. See our RDIMM vs UDIMM page for the buffering distinction on its own.

Registered ECC: the server-platform default

Most rack and tower servers specify RDIMM that also carries ECC - the register chip that buffers command/address signals, combined with the extra correction bits, on one module. This is the class Samsung's own current DDR5 RDIMM lineup lists at up to 256GB per module, and it is what the phrase "registered ECC" is normally shorthand for on a server spec sheet.

DDR5 on-die ECC: a different layer, not a replacement

Lenovo Press describes DDR5's on-die mechanism directly: "DDR5 introduces an automatic on-die single-bit error correction (SEC) that fixes possible errors before outputting data to the controller. Error correction continues to the system level, where each DDR5 channel consists of 32 data bits and 8 ECC (error checking and correction) bits."

Read carefully, that passage describes two separate correction points inside the same channel, not one replacing the other: on-die SEC corrects errors inside the DRAM chip itself, before the data ever leaves the package, while the module's own 8 ECC bits per 32 data bits correct errors that can occur afterward - in transmission across the channel to the controller. A DDR5 module still has to be an ECC-class module (ECC UDIMM or registered ECC) for that second, system-level layer to exist at all; on-die correction is present on DDR5 more broadly and does not by itself make a non-ECC DDR5 module equivalent to an ECC one. We have not found a vendor source that states this using the words "not a substitute" specifically, so we are describing the mechanism as documented rather than attributing that phrase to Lenovo or anyone else.

How to tell which one you actually have

Read the module's own printed label or the listing's stated specification first - "ECC," "Registered," and "Unbuffered" are normally spelled out directly rather than left for a buyer to infer. Where a part number is all that's available, manufacturer-specific grammar can confirm it independently: Micron encodes ECC via a 72-bit data width segment in its own part-numbering scheme, which is how our part-number pages verify RAM listings without guessing. See also our how-to-check-installed-memory page for reading back what is physically installed on a running system.

The OEM's own list is the final authority

Which ECC class (if any) your specific server requires is set by the platform, not by this page. See our OEM-verified platform pages for models we have checked directly, or the manufacturer's own memory-configuration guide for anything we have not yet built.

Live, split by ECC mechanism

Cheapest in-stock ECC UDIMM and registered (RDIMM/LRDIMM) modules right now, computed live.

ClassBrandModuleCapacity$/GB
RDIMMSK HynixHMAA8GL7MMR4N64GB$4.61Buy Now
RDIMMA-TechM393A2G40DB0-CPB16GB$5.44Buy Now
RDIMMSamsungM393A8K40B22-CWD64GB$6.23Buy Now
LRDIMMSupermicroB07DFLTWN364GB$6.23Buy Now
RDIMMSamsungM393A2G40EB1-CPB16GB$7.12Buy Now
RDIMMMICRONMTA36ASF4G72PZ32GB$7.19Buy Now

Frequently asked questions

What's the difference between ECC and non-ECC memory?

Non-ECC memory carries only data bits and cannot detect or correct a bit flip. Lenovo Press states the DDR5 comparison directly: "Non-ECC DIMMs have only the 32 data bits per channel, and will only be used in end-user devices." ECC memory adds extra bits specifically to detect, and in most cases correct, single-bit errors before they corrupt data the system relies on.

Is ECC UDIMM the same as RDIMM?

No - they answer different questions. ECC UDIMM adds error-correction bits to an otherwise unbuffered module; it does not add the register buffer chip that defines an RDIMM. A server that requires registered memory will not accept ECC UDIMM as a substitute - see our RDIMM vs UDIMM page for the buffering distinction itself.

What does "registered ECC" mean?

It describes an RDIMM that also carries ECC - the register buffer chip (which reduces electrical load on the memory controller, letting more ranks share a channel) plus the extra ECC bits for error correction, combined on one module. This is the default class most rack and tower servers specify, distinct from a UDIMM that happens to also carry ECC.

Does DDR5's on-die ECC mean I don't need ECC memory anymore?

No - on-die ECC and system-level (registered) ECC are not the same layer, and one is not a substitute for the other. Lenovo Press describes DDR5's on-die correction as fixing "possible errors before outputting data to the controller," after which "error correction continues to the system level, where each DDR5 channel consists of 32 data bits and 8 ECC... bits." That is two separate correction points in the same channel, not one mechanism replacing the other - a DDR5 module still needs to be an ECC-class module (ECC UDIMM or registered ECC) for the system-level correction Lenovo describes to exist at all.

Is on-die ECC unique to DDR5?

As documented, yes for this generation's automatic per-chip single-bit correction - Lenovo Press specifically introduces it as a DDR5 feature. We have not sourced an equivalent DDR4 on-die mechanism from any vendor this session, so we are not claiming DDR4 has one.

How do I know if a specific module is ECC or non-ECC?

Read the module's own label or the listing's stated specification - "ECC" is normally spelled out directly, and our own part-number decoder confirms it independently on the major brands we track (Micron, for example, encodes ECC via a 72-bit data width in its part-numbering grammar). Never assume ECC from a server chassis alone; check the module itself.

Sources

Specification claims on this page were checked against the pages below on 16 September 2026. Where a widely-repeated figure did not survive that check, we report what the source says now and say so in the text rather than repeating the familiar number.

  1. Lenovo Press (André Rossouw) - Introduction to DDR5 Memory (LP1618) (lenovopress.lenovo.com, 2026-09-16)
    Used for: Published 10 Nov 2022. Quoted verbatim (on-die ECC): "DDR5 introduces an automatic on-die single-bit error correction (SEC) that fixes possible errors before outputting data to the controller. Error correction continues to the system level, where each DDR5 channel consists of 32 data bits and 8 ECC (error checking and correction) bits." Quoted verbatim (RDIMM/UDIMM, non-mixable): "RDIMMs (registered DIMMs), will use a 12 V input to the voltage regulator, whereas UDIMMs (unregistered DIMMs) will use 5 V. RDIMMs and UDIMMs cannot be mixed in a system." Quoted verbatim (non-ECC): "Non-ECC DIMMs have only the 32 data bits per channel, and will only be used in end-user devices."
  2. Lenovo Press (David Watts) - Lenovo ThinkSystem Memory Summary (LP1021) (lenovopress.lenovo.com, 2026-09-16)
    Used for: Updated 29 Jul 2025. Real shipping-part rank notation, quoted verbatim from the ordering tables: "ThinkSystem 32GB TruDDR5 6400MHz (1Rx4) RDIMM", "(2Rx4) RDIMM", "(1Rx8) RDIMM", "(2Rx8) RDIMM", "ThinkSystem 256GB TruDDR5 6400MHz (4Rx4) 3DS RDIMM", "(8Rx4) 3DS RDIMM". On x4-class width and RAS, quoted verbatim: "9x4 RDIMMs... support lower fault-tolerance characteristics. Standard RDIMMs and 3DS RDIMMs support two 40-bit subchannels (that is, a total of 80 bits), whereas 9x4 RDIMMs support two 36-bit subchannels (a total of 72 bits)... The extra bits in the subchannels allow standard RDIMMs and 3DS RDIMMs to support Single Device Data Correction (SDDC), however 9x4 RDIMMs do not support SDDC."
  3. Kingston Technology - Different Types of Server Memory and How They Impact Performance (kingston.com, 2026-09-14)
    Used for: Quoted verbatim: "Different memory types (UDIMM, RDIMM, LRDIMM, MRDIMM): Cannot be mixed in the same system. If mixed, the server may not boot."
  4. Samsung Semiconductor - RDIMM module family page (spec boxes + buyer FAQ) (semiconductor.samsung.com, 2026-09-16)
    Used for: Spec boxes (capacity ceilings, current lineup): DDR5 RDIMM up to 256GB (1.1V, up to 8,000 Mbps); DDR5 MRDIMM up to 256GB (1.1V, up to 8,800 Mbps); DDR5 UDIMM up to 48GB; DDR5 ECC UDIMM up to 32GB; DDR5 SODIMM/ECC SODIMM up to 32GB; DDR4 RDIMM up to 256GB (1.2V, up to 3,200 Mbps); DDR4 UDIMM up to 32GB; DDR4 SODIMM up to 32GB. No DDR4 LRDIMM is listed in Samsung's current lineup on this page. Platform-fixed FAQ, quoted verbatim: "Different DDR generations, such as DDR4 and DDR5, cannot be mixed because they use different slot designs, signaling methods, and electrical characteristics, making them incompatible with each other." Same-generation speed-mixing FAQ, quoted verbatim: "For memory modules within the same DDR generation but with different speeds, they can be mixed, but all modules will operate at the speed of the slowest one... If it attempted to run at a higher speed than one module can handle, the system would become unstable or fail to boot."