Choose your server model, say how much memory you want, and the finder returns DIMM configurations that fit the board's slots, speed and per-module limits — priced from in-stock modules we track today. Boards we do not list get a generic DDR4/DDR5 path. Your vendor's QVL is still the final word on an exact part number.
Operating speed depends on CPU generation and DIMMs per channel. Always verify against your vendor's QVL for your exact configuration.
Pick a server model or a generic platform above to see configurations.
Every configuration above survives four checks, and it is worth stating them plainly because the value of a tool like this rests entirely on whether you can audit its reasoning. There is no proprietary scoring, no weighting, no machine-learned notion of a good match — there are four rules taken from the OEM manual for that specific server, and a module either satisfies them or it does not appear.
The first check is generation. A DDR4 board takes DDR4 and a DDR5 board takes DDR5, and there is no adapter, no BIOS setting and no compatibility mode that changes this. The keying is physically different. This check is trivial and it is also the one that catches the largest share of bad purchases, because generation is the detail buyers most often assume rather than verify when they are shopping by capacity and price.
The second is module type. Registered, load-reduced, unbuffered ECC and 3DS stacked modules are electrically different propositions, and a platform documents which of them its memory controller drives. We read that list from the manual and apply it literally. Where a manual permits UDIMM-ECC we allow UDIMM-ECC; where it is silent, we do not infer permission from the fact that a similar board in the same family allows it. Silence is treated as exclusion, which occasionally makes the finder more conservative than your board really is — a trade we accept, because the alternative failure mode is recommending memory that will not post.
The third is speed, and it is asymmetric in a way that surprises people. A module faster than the platform simply runs slower: the memory controller clocks the whole bank down to what it supports, and the configuration is valid but wasteful. We show those, flagged with the speed they will actually run at, because sometimes the faster module is the cheaper one and buying it is the right call. A module slower than the platform's documented minimum is a different case entirely — that is not a downclock, it is a part the board does not support — and those are excluded rather than flagged.
The fourth is capacity, applied at two levels. Where the OEM states a maximum per DIMM we enforce it, so a 256GB module never appears for a board documented to 128GB. And where the OEM states a maximum total, no configuration is generated that exceeds it, however many free slots the arithmetic suggests. Slot count is a hard ceiling in the same way: the finder will not propose sixteen modules for an eight-slot board, and in an add-to-what-I-have path it only uses the slots you told us are empty. It never quietly assumes you will pull a DIMM you already own — if removing modules is on the table, that is the replace path, and you choose it.
What the finder deliberately does not do is invent a fallback. If no in-stock module chain reaches your target on your platform, it says so and points you at the full compatible-module list and the shortage tracker. A tool that always produces an answer is a tool that will eventually produce a wrong one.
Results are ranked by total price because that is the number you actually pay, but price rank and buying advice are not the same thing. The configuration at the top of the list is very often the one that fills the most slots with the smallest modules, because small DDR4 modules carry the lowest cost per gigabyte in almost every market condition we have tracked. It is a genuinely good answer if the server has reached its final capacity and will never be revisited.
It is a poor answer if the machine has a future. Filling all twenty-four slots with 16GB modules to reach 384GB means the next upgrade starts by throwing away everything you just bought, because there is no empty slot left to grow into. Reaching the same 384GB with twelve 32GB modules costs more today and leaves twelve slots open, which is the difference between an incremental upgrade later and a full replacement later. The slots-free and headroom figures on each result exist so this trade is visible at the moment of decision rather than discovered eighteen months afterwards.
There is a second, quieter cost to the fill-every-slot approach. On many platforms, populating every channel with the maximum number of DIMMs per channel reduces the supported memory speed — the manual's speed table is a function of population, not just of the modules. Our speed figures come from the manual's documented range, and where the OEM attaches a caveat we surface it in the note above the results. This is a large part of why the QVL disclaimer above the configurations is not boilerplate: population interacts with speed in ways that depend on your exact CPU.
The same fewer-bigger-modules logic that governs drive upgrades applies here, and if you want to reason about it in the storage context first, the upgrade planner models exactly that trade for disks: whether to fill the bays you have with cheap capacity or buy fewer, larger units and keep room to grow.
Server memory has an unusually forgiving second-hand market, and the reason is technical rather than sentimental. DRAM does not wear out the way flash does — there is no write-endurance budget being consumed — and ECC modules report correctable errors long before they become uncorrectable ones. A used registered DIMM that posts, passes a memory test and reports clean through a few weeks of ECC logging is, for practical purposes, a working module. That is a very different risk profile from a used SSD with an unknown number of program-erase cycles behind it.
The practical split follows generation. DDR4 pulls from decommissioned fleets are abundant and cheap, and for a homelab or a non-critical tier they are usually the correct purchase. DDR5 is the opposite: the servers that would supply a used market are still in production service, so what looks like a bargain DDR5 RDIMM listing deserves more scepticism than a DDR4 one. The finder does not distinguish between new and used listings when ranking — it ranks on price and in-stock status — so the condition of a specific listing is something to check on the module page before you buy.
Whichever you choose, read the warranty terms rather than the warranty length. A seller warranty and a manufacturer warranty are different products, and the gap between them is where most disappointment lives — our refurbished warranty guide covers what each actually obliges, and the shortage tracker covers which capacities are hard to source at any condition right now.
Usually yes, and it runs at the platform's speed rather than its own. A DDR4-3200 module in a board that tops out at 2933 MT/s operates at 2933 — you paid for headroom you cannot use, but the module works. The finder shows this explicitly as a downclock note on any configuration where the module is faster than the platform ceiling. The reverse is not true and is why the finder excludes it: a module slower than the platform's minimum is not a downclock case, it is an unsupported one.
Within a channel, no — matched sets are the rule, same capacity, rank and speed. Across channels some platforms tolerate it and some refuse to boot. This is exactly the kind of per-model behaviour we will not guess at: where the OEM manual states a population rule, the finder prints that rule verbatim under each configuration; where it does not, you get the generic matched-sets line and a pointer to your manual. Every configuration the finder generates is a uniform population precisely so it is valid everywhere.
Use the generic platform path. Pick your memory generation (DDR4 or DDR5) and module type (RDIMM or UDIMM-ECC) and the finder applies platform constraints only — it will show you what the market has at each capacity and what a given DIMM count costs, without pretending to know your slot count or maximum memory. Those two numbers come from your board manual, and your vendor's QVL remains the authority on the exact part number.
For DDR4 the used and pulled market is deep, priced well below new, and the failure mode of a bad module is immediate and obvious rather than gradual. For DDR5 there is effectively no refurbished market yet, because the fleets that would feed it are still in service — you are buying new whether you like it or not. Either way check what the warranty actually covers before you buy; our refurbished warranty guide walks through the difference between a seller warranty and a manufacturer one.
Every price is our own tracking of Amazon US listings, refreshed every 30 minutes, and the configurations are computed from that data when you load this page rather than from a cached list. The freshness line under the heading shows the newest RAM price timestamp we hold. Out-of-stock modules are excluded from configurations entirely — a configuration you cannot buy today is not a recommendation.
RDIMM (registered) puts a register between the memory controller and the DRAM chips, which lets a platform address far more modules and far more capacity at the cost of a small latency penalty. UDIMM-ECC (unbuffered ECC) skips the register: lower latency, error correction, but much tighter capacity limits. They are not interchangeable — a board expects one or the other, and the finder only ever offers the types the OEM documents for that platform.
No, and it is worth being blunt about that. The finder applies conservative criteria drawn from the OEM manual — generation, permitted module types, minimum supported speed, and the per-DIMM maximum where the manual states one. A QVL is narrower: the vendor validates specific part numbers against specific CPU and population combinations. Use the finder to establish what compliant memory actually costs today; use the QVL to bless the exact part before a production order.