Buying Pulled Server Memory: Grades, What to Verify, and When Not to Buy It

Short answer: pulled server memory is the best value in the market when you know what the grade promises and you test before production. It is an expensive lesson when you buy untested modules for a platform that trains memory timings aggressively. The price difference between grades is a statement about who carries the risk of a bad module — nothing more, and nothing less.

What each grade actually promises

Grade names are used loosely across the market, so here is what each one means on our quotations — and it is restated on the quote line itself so it is not left to interpretation.

The distinction between refurbished and pulled-tested is the one buyers most often ask about. They both mean verified working. Refurbished emphasises inspection and presentation — the module has been cleaned and gone through a bench check. Pulled-tested emphasises provenance and function — it came out of working equipment and it starts and passes. If what you care about is that it works, either will do. If what you care about is how the label and the contacts look when it arrives, ask for refurbished.

Grade applies per line, not per order. A single order can contain a new line and a pulled-tested line, and each is quoted and checked separately.

Why the price gap exists

Pulled-untested modules are cheaper for exactly one reason: nobody has established that they work. That is a real saving when you have a test bench and a known-good host, because you are buying the labour back. It is a false saving when you do not, because the cost of the failure is not the module — it is the evening you spend isolating which of twelve modules will not train, on a machine you needed working.

Do the arithmetic honestly before choosing the cheap grade:

• If you have a spare host and a memory test you trust, untested is a legitimate buy.
• If the modules are going straight into a revenue machine, tested is the cheaper grade once you price your own time.
• If you are buying for resale and do not test, you are passing the risk to your own customer, not removing it.

The three checks that actually matter

When a lot arrives — or before you pay, if you are inspecting in Shenzhen — these are the checks that catch real problems, in order of usefulness.

1. Match the label against the quotation

Read the part number physically printed on the module label and compare it with the line you were quoted. Not the box, not the invoice — the module. Capacity, rank and organisation are printed together, for example 16GB 1Rx4 PC4-23400. Server vendors also print their own service part numbers, which is why a single module can legally carry several identifiers: our 16GB DDR4-2933 1Rx4 line cross-references DELL AB687113, HP P00920-B21, HP P06187-001 and HP P03051-091, and our 32GB 2Rx4 line cross-references DELL SNP8WKDYC, DELL 8WKDY, HP P03052-09S and HP P12981-091. These are interchangeable server-vendor part numbers, not an affiliation claim — any of them is the same memory, and all of them should be accepted as a match for that line.

2. Confirm what the module reports, not what it says

Install one module in a known-good host and read the running configuration. A module labelled 2Rx4 that reports as 1Rx4 is not the part you bought, and the reason it matters is that rank and organisation determine how many ranks the channel can carry — the thing that decides whether your intended population works at all. Checking capacity alone will not catch this.

3. Test in the population you intend to run

One module passing on a bench tells you the module is alive. It does not tell you that twelve of them will train together at the speed grade you expect. Where you can, populate a host the way the production machine will be populated and run a full memory test before committing the lot. This is the check that catches the failures that only appear at scale — a single channel training down, or an initialisation failure that depends on the order modules are fitted.

Which platforms punish used modules

Pulled memory is not equally suited to every machine. The risk concentrates in three places:

• Firmware that validates module part numbers. Some OEM servers refuse to boot on modules whose SPD data is not on an approved list. The module can be perfectly good and still not post in that specific machine.
• Platforms that train timings aggressively. The tighter the memory training, the more a mixed population of modules from different lots can cost you in stability — usually presenting as intermittent errors under load rather than a clean failure at boot.
• Machines under a support contract. Fitting non-OEM or non-approved memory can invalidate support on some platforms. Check the terms before you buy, not after.

Where used memory works best: general-purpose servers you control, lab and test machines, GPU host systems where you are already mixing components, and platforms with a permissive memory QVL. Ask us and we will tell you what we know about your specific platform before you commit — including when the answer is that you should not buy used for it.

Mixing modules: what works and what does not

• Module classes never mix. RDIMM, LRDIMM, 3DS RDIMM and 3DS LRDIMM are separate classes, and mainstream DDR4 platforms halt during memory initialisation if more than one class is populated — in the same channel or across sockets.
• Different lots of the same class usually train down. The platform takes the slowest common set of timings across the channel. It will boot; you may not get the speed grade you paid for.
• Different capacities per channel are the compromise case. Supported at reduced capacity in most configurations, but if the result has to be reproducible, buy the whole population as one matched set.

When new is the only sensible choice

We would rather say this than sell you the cheaper grade.

• The machine is revenue-critical and cannot tolerate a return.
• The platform is under a support contract that requires OEM parts.
• The modules are going into a validated appliance where the vendor owns the bill for any fault.
• The price gap has narrowed to the point where the risk is not worth taking — this happens in older DDR generations where new stock is scarce and both grades converge in price.

If any of those describe your order, say so and we will quote the new part. We will also tell you when the used grade we hold is not a good fit and you should be buying elsewhere.

What to ask for with your quotation

Whatever grade you buy, these should come as standard, not as a favour:

• Lot photographs of the actual modules, showing labels and contacts.
• A test report for tested and refurbished grades, showing what was run and what it returned.
• The condition grade restated on the quote line, so the delivered goods can be checked against what was quoted.
• The OEM cross-reference for the line, so you can match against whichever part number your own records use.

Inspection in Shenzhen is welcome before shipment if you are local or visiting.

Send us the part number you have

Send whichever number is on your own records — our SKU, the OEM service part number, or just the specification. We will match it against physical stock, tell you the grade we hold, and quote in writing with lot photographs and a test report. If used is the wrong buy for your platform, we will say so.

Ask Ms Aya about availability and grades, or tell us the part you need if it is not listed.