ECC UDIMM has always cost more than plain desktop memory. In the 2026 memory shortage it costs a great deal more. A single 32GB ECC stick that a homelab builder could have bought for well under two hundred dollars two years ago now clears eight hundred, and many of the cheap kits that surface in a search are the wrong module for the board they will be dropped into. Buying well right now is less about finding ECC and more about not overpaying for the wrong part.
This guide assumes the ECC decision is already made. Whether a homelab actually needs ECC, and whether a given board will switch it on and report it, is a separate question answered in the companion piece on ECC RAM for a homelab. What follows is the shortlist of unbuffered ECC modules worth buying today, the capacity to pick per slot, and how to match one to the machine without paying registered-DIMM money for a part a workstation board will refuse to post. If error correction turns out not to be a hard requirement, the broader DDR5 memory shortlist for a homelab and workstation covers the non-ECC options.
One caveat runs through every price here. In the shortage, DDR5 kit prices have moved week to week, so each figure below is a band to sanity check against the live listing, not a number to quote back. The specifications and prices were verified against manufacturer datasheets and current listings in September 2026. Every pick is a 288-pin unbuffered ECC module, which is what workstation and entry-server boards take. Registered modules, the kind that go in a dual-socket server, are out of scope and are keyed so they will not seat in these boards anyway.
How the four kits compare
| Module | Capacity | Generation, speed | Rank | ECC width | Best for | Live band |
|---|---|---|---|---|---|---|
| A-Tech DDR5 ECC UDIMM | 16GB | DDR5-5600 | 1Rx8 | 72-bit | cheapest way onto a DDR5 ECC board | ~$460 to $500 |
| Samsung M324R | 32GB | DDR5-4800 (JEDEC) | 2Rx8 | 72-bit | the module a board’s tested list usually names | ~$850 to $920 |
| Kingston Server Premier | 48GB | DDR5-5600 | 2Rx8 | 72-bit | most memory per slot, brand plus lifetime warranty | ~$1,650 to $1,850 |
| OWC DDR4 ECC UDIMM | 32GB | DDR4-3200 | 2Rx8 | 72-bit | a DDR4 ECC board you already own | ~$280 to $330 |
Every module in the table carries a true 72-bit width, which is 64 bits of data plus 8 bits of ECC, arranged on a DDR5 module as four ECC bits per 32-bit sub-channel. That extra width, not the on-die correction that ships on all DDR5, is what lets the memory controller detect and correct errors across the whole path and report them to the operating system. It is a lighter correction scheme than a registered server module carries, but it covers the transit path between the module and the controller that on-die correction leaves exposed. The difference between on-die and side-band ECC, and why one is not a substitute for the other, is covered in the companion guide. Here it is enough to say the width is the thing you are paying for.
1. A-Tech 16GB DDR5 ECC UDIMM: the cheapest way onto a DDR5 ECC board
The cheapest genuine DDR5 ECC UDIMM from a named module builder is a single 16GB stick. It is the part to reach for when the goal is to confirm a new board actually enables and reports ECC before committing to a full set, or to seat a modest amount of correcting memory in a firewall, a small hypervisor, or a backup target that does not need much.
The specification is standard JEDEC: DDR5-5600, PC5-44800, single rank as 1Rx8, 1.1V, 288-pin, with the 72-bit width that makes it real ECC. It runs at its rated JEDEC speed on a supporting board with no overclock profile to enable, which is exactly what you want in a machine bought for uptime. Expect it to land around $460 to $500, so check the live price before ordering, because a single 16GB stick at that money is only sensible as a starter or a light build.
Skip it if you are building anything memory-hungry. Two of these to reach 32GB costs more than a single 32GB stick and burns two slots, which matters on the four-slot boards most of these CPUs use. As a first stick to prove the platform, or as the whole memory budget of a small always-on box, it earns its place. As the foundation of a workstation, it does not.
2. Samsung 32GB DDR5 ECC UDIMM: the module your board was tested with
When a workstation board publishes a memory support list, the parts on it are overwhelmingly first-party modules from the companies that make the DRAM. The Samsung M324R 32GB module is one of those. It is the default 32GB stick to buy when you want the module the board vendor actually validated, rather than an equivalent assembled from the same chips by a third party.
It is a dual-rank 2Rx8 module at DDR5-4800, PC5-38400, CL40, 1.1V, 288-pin, 72-bit. The 4800 rating is the JEDEC baseline for ECC UDIMM rather than a fast bin, and that is deliberate: correcting memory is sold to run at standard speed on a supporting platform, not to chase a benchmark. The live price sits around $850 to $920, so confirm the current number before you buy a pair.
Pass on it when your board’s support list specifically calls for a 5600 module and you would rather match that number than the JEDEC baseline. In that case a dual-rank A-Tech 32GB DDR5-5600 ECC UDIMM runs a similar price, around $840 to $890, and clocks a step higher, at the cost of not being the exact first-party part on the list. For most builds the Samsung is the safer default; the A-Tech is the value alternative when you want 5600 and trust the assembler tier, which the buying section below explains.
3. Kingston Server Premier 48GB DDR5 ECC UDIMM: the most memory per slot
Four slots is the common ceiling on these boards, and the non-binary 48GB module is how you push total capacity without adding sticks. Two of them reach 96GB in two slots, leaving room to grow; four reach 192GB, which is about as much as an unbuffered ECC platform will carry. The Kingston Server Premier part is the one to buy when capacity and a real warranty both matter.
Kingston builds this part on Hynix M-die as a dual-rank 2Rx8 module at DDR5-5600, CL46, 1.1V, 288-pin, 72-bit, and programs its timings to the JEDEC standard with no overclock profile. The Server Premier line is platform validated with both Intel and AMD and carries a lifetime warranty, which is the practical reason to pay for the brand rather than an assembled equivalent. Plan for roughly $1,650 to $1,850 per stick, and check the live price, because 48GB modules are where the shortage bites hardest.
The case against it is narrow. If a board’s support list caps module size at 32GB, which some workstation boards do, a 48GB stick may run there but is not guaranteed to, and at this price you do not want to gamble. If the capacity is what you are after but the brand premium is not, an A-Tech 48GB DDR5-5600 ECC UDIMM covers the same slot for around $1,250 to $1,350, trading the lifetime warranty and QVL match for a lower number.
4. OWC 32GB DDR4 ECC UDIMM: for the board you already own
Most ECC-capable boards in service today are still DDR4. A large share of homelab ECC lives on the previous generation of workstation and entry-server platforms, and for those machines a DDR5 module is useless. The OWC 32GB DDR4 part is the clean, brand-backed way to add or expand ECC on a board you already have, at a fraction of what the DDR5 parts cost.
It is a dual-rank 2Rx8 module at DDR4-3200, PC4-25600, CL22, 1.2V, 288-pin, 72-bit, which is the JEDEC baseline for DDR4 ECC UDIMM. OWC is a workstation and NAS memory specialist rather than a fab, but the part is properly labeled, warrantied, and priced like DDR4 should be in this market, around $280 to $330. It is the one pick here that a NAS or a mailserver on aging hardware can actually justify, so check the live price and buy for the board you have.
There is only one reason to pass: building new. Nobody should buy a DDR4 board in 2026 to save on memory, and if the platform is already DDR5 this part will not seat. For an existing DDR4 workstation, though, it is the most sensible purchase on this page. A kit of two 16GB DDR4 ECC modules from an assembler such as NEMIX is the alternative when you want to fill two slots at once, at around $320 to $360.
How much to buy, and how many sticks
These platforms run dual channel, so memory goes in as two matched sticks, not one. A single stick works but leaves half the bandwidth on the table. The practical unit of purchase is therefore two identical modules: two 32GB parts for 64GB, two 48GB parts for 96GB, and so on. Buy them as the same part number from the same source, because mixing ranks or bins across a channel is the fastest way to a board that trains at a lower speed or refuses to post.
Prefer fewer large sticks over more small ones. Populating all four slots, two modules per channel, forces the memory controller to a lower clock than running two slots does; the drop is real and it is imposed by the CPU, not the modules. Two 48GB sticks will hold their rated speed where four 24GB sticks would not, and they leave two slots open for a later upgrade. The one time to fill all four is when you have already maxed the largest module the board supports and simply need more total capacity than two slots can hold. If the target is a ZFS box, size the total against the pool rather than a rule of thumb; the ZFS RAID levels you choose set how much the ARC can usefully cache.
Brand tiers, and when an assembler module is fine
ECC UDIMM comes from two kinds of vendor. The first tier is the companies that make the DRAM and assemble the modules themselves: Samsung, and Kingston building on Hynix or Micron die with published validation. Their parts are the ones that appear on board support lists by exact part number, and they carry the manufacturer’s own warranty. When a board vendor’s list names a specific module, buying that module removes the compatibility question entirely.
The second tier is the assemblers: A-Tech, NEMIX, OWC, and similar. They buy the same DRAM chips and build modules to the JEDEC spec, usually with a lifetime replacement warranty, at a lower price. For a homelab that is running standard JEDEC speeds anyway, an assembler module is a reasonable buy, and in this shortage it is often the only part in stock at a given capacity. The trade you accept is that it will not be the exact SKU on a QVL, so if a board is fussy the burden of proving compatibility falls on you. The rule of thumb: match a first-party module to a QVL when the board publishes one and the part is available; use an assembler when it is not, or when the saving is large and the platform is known to be tolerant.
Buying ECC UDIMM in the 2026 shortage
The prices on this page are two to six times what the same parts cost before the shortage, and they are not stable. Stock is thin, listings flip to scarce or to a longer lead time without warning, and some enterprise-leaning SKUs show a price only to a business account. That reshapes the buying advice in a few concrete ways, laid out against the wider rise in RAM and storage prices.
Buy the capacity you need now rather than staging upgrades, because the next stick may cost more, not less, and matched pairs bought months apart may no longer match. Do not pay registered-DIMM money for an unbuffered part; if a 32GB module is priced like a server RDIMM, it is either mispriced or the wrong module, and it will not help a workstation board. Check the exact part number against the listing, not just the capacity and speed, because ECC and non-ECC sticks of the same size sit side by side in search results at similar prices. And weigh the DDR4 option honestly: if the workload runs on an existing DDR4 ECC board, staying there through the shortage is far cheaper than a new DDR5 platform bought at today’s memory prices.
Confirm the board does ECC before the money leaves
The single most expensive mistake with ECC memory is buying it for a machine that quietly ignores it. Mainstream consumer boards do not enable ECC UDIMM at all. On the Intel side, unbuffered ECC on a desktop socket runs through the W680 workstation chipset, not the consumer Z, B, or H boards. On the AMD side, the Ryzen memory controller supports ECC, but whether an AM5 board actually enables and reports it is a firmware decision: it is validated on the Ryzen Pro line and left to the board vendor on the rest, with some boards reporting corrected errors fully and others not at all.
That matrix, which board and CPU combinations actually turn ECC on and report corrected errors, is the whole subject of the companion guide on whether a homelab needs ECC, alongside the ZFS storage server build and how much memory a Proxmox homelab needs. Read the platform section there before spending eight hundred dollars a stick. The modules on this page are correct; they only do their job on a board that agrees to let them.
What to check on the sticks the day they arrive
Assembler listings and shortage-era stock both raise the odds of receiving the wrong part, so the first thing to do with new ECC memory is confirm the board sees it as ECC before trusting it. On Linux the acceptance check is one line against the firmware tables.
sudo dmidecode -t memory | grep -iE "Total Width|Data Width|Error Correction"A genuine ECC UDIMM on an ECC-capable board reports a Total Width of 72 bits against a Data Width of 64 bits, and an Error Correction Type of Single-bit ECC on the memory array. If the total and data widths match, or the correction type reads None, the board is treating the module as plain memory, and the extra width you paid for is doing nothing. Confirming that the correction is not only present but actively catching errors, through the kernel’s EDAC counters and rasdaemon, is the operational check described in the companion guide; the widths above are the one to run the day the parts land, before the machine goes into service.
