Mixing Memory Kits, and XMP

Two kits with the same part number are not a matched kit, and XMP is an overclock rather than a default. Those two facts explain most of the memory problems people hit, and both are easy to miss because the packaging implies otherwise.

A note on this guide: this rests on how memory is validated and specified, not on kits tested here. This page has no product recommendations — memory is a commodity where the specification is the product. See how we test.

The problem, as it appears

A machine that has been fine for two years gets more memory. Same brand, same speed, same part number as the pair already in it.

Now it crashes occasionally, or fails to start at all, or runs at a speed lower than either kit is rated for. Nothing is faulty, and every component is exactly what the box said.

Why "identical" disappoints

Because a kit is validated as a set, and validation is the thing you bought.

When a manufacturer sells two modules as a kit, they have been tested together at their rated settings. Two kits of the same part number were each validated internally, and no one has ever tested all four as a group — so whether they cooperate at the rated profile is something you find out rather than something anyone promised.

Silicon varies between production runs, and a part number does not pin down which run you received. This is why "buy the capacity you want in one kit" is the standing advice: not because mixing always fails, but because the guarantee only exists inside a kit.

Populating four slots is also harder on the processor's memory controller than two, which is a separate reason a machine can be stable with two modules and marginal with four at the same settings.

What actually matters

XMP is a profile, not a default

This is the part worth internalising.

Memory has a standard speed it is guaranteed to run at. XMP — and its equivalents under other names — is a set of faster speeds, tighter timings and higher voltages stored on the module, which the board can apply in one step.

Those settings are above the guaranteed standard. The manufacturer has tested them; your particular board and processor have not agreed to anything. It usually works, which is why it is normal practice and why the language around it treats it as switching the memory on properly.

But the consequence matters: if a machine becomes unstable after enabling the profile, the profile is a suspect. People rule it out because it came with the memory, and that is exactly backwards.

The speed you bought is often not the speed you get

Boards publish supported memory speeds by configuration, and the supported figure typically drops as you add modules. Four modules at a high rated speed is a demanding configuration, and the board may run them slower or refuse the profile entirely.

That is not a fault. It is the board declining an overclock it cannot hold in that configuration, and the honest response is to accept the lower speed rather than to force it.

Stability is a week, not a boot

The same standard that applies to undervolting applies here, for the same reason: a machine that boots and passes a memory test can still fail on a particular workload days later.

Run it normally for a week before concluding a configuration is good. If anything odd appears, the profile is the first thing to turn off — and undervolting to cut fan speed makes the same argument about judging stability by use rather than by a passed test.

Slots are specified and it is not the obvious pair

For two modules, boards specify which two slots to use, and it is usually not the pair closest to the processor. The manual has a table for this.

Getting it wrong costs bandwidth if the machine runs, and stops it starting if it does not — which is step two of a PC that will not POST for exactly this reason.

More capacity is not more speed

Adding memory helps a machine that was running out of it. Beyond that point, extra capacity changes nothing you can measure.

Check what is actually in use under the load you care about before buying. It is a common upgrade bought on the general principle that more is better, and in this specific case the principle stops applying at a threshold you can observe.

What to do, in order

1. Buy one kit at the capacity you want. This avoids the entire problem and is almost always available.
2. If you must add to existing memory, expect to run at the slower of the two rated speeds, or at the standard speed with no profile.
3. Enable the profile, then test properly — a week of ordinary use, not a boot and a benchmark.
4. If anything is unstable, turn the profile off first. Confirm the modules are fine at standard speed before suspecting anything else.
5. Accept a slower stable configuration. The performance difference between the profile and the standard speed is small; the difference between stable and not is not.

The mistake to avoid

Treating a crash after a memory upgrade as a faulty module. The modules are usually fine. What has changed is that a configuration nobody validated is being asked to run at settings nobody promised, and the first diagnostic step is free: turn the profile off and see whether the machine settles.

The second mistake is buying more memory to fix a machine that is slow for another reason. Capacity helps only when capacity was the constraint, and a machine that is slow because of a full drive, a thermal problem or a background process will be exactly as slow with twice the memory — the ordered version of that diagnosis is in what goes wrong with a gaming PC after year one.