✅ Updated Nov 2026
Affiliate Disclosure: As an Amazon Associate, we earn from qualifying purchases. Links marked "Check on Amazon" are affiliate links — learn more.

Memory is the one part of a gaming PC where the box speed and the running speed are often two different things. A kit sold as DDR5-6000 boots at a slower standard speed until someone enables its profile in firmware, and no operating system, Linux included, can change that after the fact. That fact shapes every RAM decision below.

This guide explains the settings side of memory: how much capacity a game library actually uses, what speed and timings mean in plain terms, where XMP and EXPO live and why Linux only sees their result, and how to prove a kit is stable. The five kits at the end are examples of each class, with maker-stated specifications.

As an Amazon Associate, we earn from qualifying purchases.

Quick answer

Buy a two-stick kit matched to your platform, 32GB for most gaming builds, and enable its XMP or EXPO profile in the motherboard firmware setup. Linux never sets memory speed itself; it simply runs at whatever the firmware trained. Check the result with dmidecode, then prove stability with a bootable memory test.

Capacity honestly, by library

Capacity is the least mysterious memory specification and the one most often oversold. A game, the operating system, a browser, and a launcher all share the same pool, and what matters is the total working set, not the number printed in one game’s requirements list.

Sixteen gigabytes still runs a large share of current games on Linux, because Linux itself is light and most titles were designed around console memory budgets. The pressure comes from everything around the game: shader compilation caches, a browser with many tabs, Discord, recording software, and open-world titles that stream assets. When the pool runs out, Linux swaps to disk, and the symptom is stutter and hitching, not an error message.

Thirty-two gigabytes is the honest default for a new gaming build. It covers heavy titles, background applications, and the page cache Linux uses to keep recently loaded game files in memory. Sixty-four gigabytes is a workstation decision, not a gaming one: it suits players who also run virtual machines, compile large projects, host game servers, or edit video. Buying it purely for games leaves most of it idle.

One structural rule matters more than the headline number: buy capacity as a matched two-stick kit. Two sticks give the memory controller an easier training job at high speeds than four sticks, and a matched kit means both modules carry identical chips and timings. Filling all four slots on current DDR5 platforms frequently forces the speed down, which trades away the very thing the fast kit was bought for.

Speed and timings in plain terms

A memory speed rating such as DDR5-6000 describes data transfers per second, in millions. Timings describe delays measured in clock cycles, and the first of them, CAS latency or CL, is the one printed largest. Neither means much alone: a cycle at a higher speed is shorter in real time.

The conversion is simple: real latency in nanoseconds equals the CL number divided by half the transfer rate, multiplied by one thousand. DDR5-6000 CL30 works out to ten nanoseconds. DDR4-3600 CL18 also works out to ten nanoseconds. DDR5-5600 CL46, the standard no-profile speed class, works out to roughly sixteen nanoseconds. That is why a modest DDR4 kit and a fast DDR5 kit can feel similar in latency-bound moments, and why chasing frequency alone disappoints.

The practical reading is this. Moving from the fallback speed to a rated profile speed is a large step, because bandwidth and latency improve together. Moving between neighbouring fast kits, say DDR5-6000 CL30 and DDR5-6400 CL32, is a small step that most players will not feel outside a benchmark graph. On AMD AM5 systems there is an additional documented wrinkle: the platform’s memory controller is happiest at or below DDR5-6000, where it can run in a one-to-one ratio with the memory clock, which is why DDR5-6000 CL30, not the fastest kit on the shelf, became the standard recommendation for that platform.

XMP, EXPO, and what Linux actually sees

XMP is Intel’s Extreme Memory Profile; EXPO is AMD’s Extended Profiles for Overclocking. Both are data stored on a small chip on each module, telling the motherboard a tested combination of speed, timings, and voltage beyond the conservative JEDEC standard every module also carries. Kits are sold in XMP-only, EXPO-only, and dual-profile versions, so the profile type is part of the product, not a detail.

Enabling a profile is a firmware act. You enter the motherboard setup during boot, select the profile, save, and the board trains the memory at those settings before any operating system loads. Linux has no tool that does this, and neither does Windows; operating-system tuner apps only monitor, they do not set the trained speed. What Linux receives is the finished result: memory already running at the profile speed, or at the JEDEC fallback if the profile was never enabled or training failed.

That fallback is the most common memory problem in new builds. The machine works, games run, and nothing complains, yet the kit sits at DDR5-4800 with loose timings because the profile was never switched on. Linux makes the state easy to read: dmidecode, run with administrator rights, reports each module’s configured memory speed, and free and /proc/meminfo show capacity and use. If dmidecode reports a lower configured speed than the kit’s rating, the fix is a reboot into firmware setup, not a Linux setting.

Stability, and the memory-test route

A profile is an overclock, even a factory one, and an overclock can be unstable in ways that look like everything except memory. On Linux the classic symptoms are a game closing without an error, a shader compile that fails randomly, a desktop session logging out, or file corruption that appears weeks later. Because these symptoms mimic driver and game bugs, memory deserves suspicion early.

The honest order of operations is to test before blaming software. The standard route is a bootable memory tester such as MemTest86+, run from a USB stick, for several full passes and ideally overnight on a new build. Errors at the profile speed mean the combination of that kit, that board, and that processor is not stable: update the motherboard firmware first, since memory training improvements arrive through firmware updates, then retest, then drop the speed one step if errors persist.

Heat plays a quiet role: modules tucked under a large air cooler with little airflow can drift into errors during long sessions that a short test missed. A ten-minute test proves very little; duration is part of the method.

How to choose

Choose by platform first. AMD AM5 boards take DDR5 only and pair best with an EXPO kit at DDR5-6000 CL30 in a two-stick 32GB configuration. Intel LGA1700 boards exist in DDR4 and DDR5 versions that are not interchangeable, so the board decides the generation before the kit does; on a DDR5 Intel board, an XMP kit at a similar speed and latency is the equivalent pick. On an older AM4 or DDR4 Intel system being refreshed, a DDR4-3600 CL18 kit is the sensible ceiling, because the platform gains nothing from faster memory.

Choose the profile type printed on the kit: an EXPO kit for an AMD build, an XMP kit for an Intel build, or a dual-profile kit if the machine may change platforms. Choose 64GB only when a named non-gaming workload needs it. And choose the plain JEDEC-class kit, the kind with no profile at all, when the machine is a home server or living-room box where stability outranks the last slice of game performance.

Who should skip this

Skip the fast-kit question if your current machine already runs a matched 32GB kit at its rated profile speed; a side-grade buys nothing you can feel. Skip EXPO and XMP tuning if the PC is a shared family machine whose crashes would cost more goodwill than a few frames are worth. Skip 64GB if gaming is the only heavy workload. And skip memory upgrades on a laptop with soldered memory or a handheld, where none of this can be changed.

G.Skill Trident Z5 Neo 32GB DDR5-6000 CL30

Check price on Amazon

The reference AM5 gaming kit: two 16GB modules with an AMD EXPO profile at the speed the platform handles best, and tight timings for its class.

  • Specs: 32GB kit, 2 x 16GB modules
  • Specs: DDR5-6000 speed rating
  • Specs: CL30-38-38-96 timings
  • Specs: AMD EXPO profile, 1.35V
  • Specs: Model F5-6000J3038F16GX2-TZ5N

Corsair Vengeance 32GB DDR5-6000 CL30

Corsair Vengeance 32GB DDR5-6000 CL30

Corsair Vengeance 32GB DDR5-6000 CL30

Check price on Amazon

The Intel-side equivalent: a two-stick XMP kit at the same speed and latency class, in a low-profile heatspreader.

  • Specs: 32GB kit, 2 x 16GB modules
  • Specs: DDR5-6000 speed rating
  • Specs: CL30-36-36-76 timings
  • Specs: Intel XMP profile, 1.40V
  • Specs: Model CMK32GX5M2B6000C30

Crucial Pro 32GB DDR5-5600

Check price on Amazon

The no-profile option: a JEDEC-standard kit that runs at its full rated speed the moment it is installed, with nothing to enable.

  • Specs: 32GB kit, 2 x 16GB modules
  • Specs: DDR5-5600 speed rating
  • Specs: CL46 timings, JEDEC profile
  • Specs: Standard voltage operation
  • Specs: Model CP2K16G56C46U5

Corsair Vengeance LPX 32GB DDR4-3600 CL18

Check price on Amazon

The DDR4 refresh pick for AM4 and DDR4 Intel boards, at the speed those platforms use best, in a short heatspreader.

  • Specs: 32GB kit, 2 x 16GB modules
  • Specs: DDR4-3600 speed rating
  • Specs: CL18-22-22-42 timings
  • Specs: XMP 2.0 profile, 1.35V
  • Specs: Model CMK32GX4M2D3600C18

G.Skill Trident Z5 Neo 64GB DDR5-6000 CL30

Check price on Amazon

The capacity pick for machines that game and work: two 32GB modules at the same AM5-friendly speed, keeping the two-stick layout.

  • Specs: 64GB kit, 2 x 32GB modules
  • Specs: DDR5-6000 speed rating
  • Specs: CL30-40-40-96 timings
  • Specs: AMD EXPO profile, 1.40V
  • Specs: Model F5-6000J3040G32GX2-TZ5N

FAQ

Can I enable XMP or EXPO from inside Linux?

No. The profile is selected and trained by the motherboard firmware before the operating system loads, and Linux only inherits the result. Use dmidecode to confirm the configured speed, then reboot into the firmware setup if the number is lower than the kit’s rating.

My kit runs at DDR5-4800 even though it is rated for 6000. What happened?

The profile was never enabled, or training at the profile failed and the board reverted to the JEDEC fallback. Enable the profile in firmware setup, update the motherboard firmware, and confirm the change with dmidecode.

Is mixing two different kits to reach more capacity a good idea?

It is the most common cause of memory that will not train at its rated speed. Modules from different kits can carry different chips and subtimings even under the same product name. Buy one matched kit instead of adding a second kit later.

How do I check how much memory a game is actually using?

The free command shows total, used, and available memory for the whole system, and /proc/meminfo carries the detailed figures. Per-process tools such as the system monitor or top show what one game holds. Available memory, not free memory, is the number that says how much room is left, because Linux uses spare capacity as file cache.

Final verdict

For a Linux gaming build, memory shopping is mostly settings discipline: a matched two-stick kit, the profile type your platform expects, capacity sized to the whole workload rather than one game, and the profile enabled in firmware. Do those four things, verify with dmidecode, and prove stability with an overnight test, and RAM becomes the part you never think about again.

Explore Our Guides & Free Tools