How to choose a desktop RAM kit: size, speed and how many sticks

Buy one box holding two sticks, 32GB in all (2 x 16GB): DDR5-6000 if your motherboard takes DDR5, DDR4-3200 if it takes DDR4, and the 5600 kit if DDR5-6000 costs more than a few percent over DDR5-5600. DDR5 means an AMD Ryzen 7000 or 9000 (the AM5 socket) or an Intel Core Ultra 200S, DDR4 means a Ryzen 5000 (AM4), and Intel 12th to 14th Gen comes on boards of either kind; look for a latency of CL30 to CL36 at DDR5-6000 and CL16 at DDR4-3200. Once it is fitted, switch on its speed profile, a setting stored on the sticks, in the machine’s setup screen, because a kit runs at a slower standard speed until you do. Two sticks beat four because AMD and Intel both rate four DDR5 sticks below two, AM5 by a third, and 2 x 8GB is a dead end because adding to it later means four sticks. This site prices kits for every desktop but has no page for your board: RAM by processor gives your chip’s memory type and rated speed, and the board maker’s manual and memory list settle the rest.

What to do in a few minutes

This is the whole job. The rest explains why, and what to do when something goes wrong.

1. Name the processor and the board. If you are building new, both names are on the boxes or the shop’s pages. On a Windows machine you already own, open PowerShell from the Start menu, and these two lines print them:

Get-CimInstance Win32_Processor | Select-Object Name
Get-CimInstance Win32_BaseBoard | Select-Object Manufacturer, Product

A prebuilt from a big maker may print only a code for its board. In that case naming the machine says where each maker keeps its memory table.

2. Look the processor up under RAM by processor, whose Desktop button narrows the list. Its page gives the memory type, rated speed, channels and ceiling. That is the chip’s limit, not the board’s.

3. Open the board’s manual at its memory section. Note DDR4 or DDR5, the number of slots, the largest stick it names and which slots to fill first. The maker’s memory support list shows which kits it has tested.

4. Pick your row in the table below. On a DDR5 row, open DDR5 desktop memory, choose New, and compare 6000 with 5600 under Speed on price per gigabyte. This site’s prices leave shipping out, so add it to your own sums.

5. Fit both sticks in the slots the manual names, usually A2 and B2, with the machine off and unplugged.

6. Start it and wait. A new DDR5 kit can leave the screen black for minutes on its first start while the board trains the memory. That is usually normal.

7. Turn the profile on using your board maker’s menu, then read the speed back in Task Manager, under Performance then Memory.

8. Test the kit while you can still return it.

The kit to buy, by platform

A platform here means the processor’s socket and the board around it. DDR5-6000 means 6,000 million transfers a second (MT/s). CL is the kit’s latency in clock cycles, and at a given speed lower is quicker. XMP and EXPO are the speed settings stored on the sticks: XMP is Intel’s name and EXPO is AMD’s.

A laptop or a mini PC is a different question, starting with whether the memory can be changed at all: can you upgrade the RAM in your laptop begins there.

The speed in that table is a rule about price. Buy RAM now or wait sets it out: DDR5-6000 at CL30 to CL36 when it costs within a few percent of DDR5-5600, otherwise the 5600 kit, and never a large premium for anything faster. On the day below, the median 6000 kit did not pass it.

These are the medians the DDR5 desktop and DDR4 desktop pages printed on November 1, 2026, with New under Condition, 16 GB under Module capacity, 32 GB under Total capacity and the speed under Speed: item prices, before shipping. DDR5-6000 sat 8.1% over DDR5-5600, so on that day’s medians the 5600 kit was the buy, and DDR5-6400 sat 14% over DDR5-6000. DDR4-3600 sat 27% over DDR4-3200, far more than the speed returns. Medians move, so read them again before you pay.

A median is not a floor, and eBay is not the only shop. Tom’s Hardware’s tracker put the cheapest 32GB DDR5-5600 and DDR5-6000 kits in US stores at the same $409 on September 14, 2026 (the DRAM shortage guide has more), so compare two kits you could actually buy, and price a store kit before you treat these medians as the market. Whether to buy at all this year is the subject of buy RAM now or wait.

Short answers

Why a kit, and why two sticks

A kit is two or four sticks that the maker tested together at the speed on the box. Corsair states the limit plainly on its support page: its kits “are only validated for their rated performance when only using the modules included in that specific kit”. Two loose sticks of the same model are not that product. Tom’s Hardware (July 31, 2026) says they “should” work and that the word carries a lot of weight, because a maker may source chips from more than one place and a later stick’s timings may not line up. A kit is binned as a set has the reasoning.

One stick against two

A desktop processor has two memory channels, two separate paths to memory, and one stick fills only one of them. A single DDR5 stick is no exception, as DDR4 vs DDR5 explains.

The 1% lows are the slowest frames, the ones you notice as stutter. Single games go further: Doom: The Dark Ages lost 17.3% on the 7600X in Tom’s test, and Marvel Rivals fell from 195 fps to 129 fps in TechSpot’s. Capacity halved as well in both tests, and both treat one stick as a stopgap. Tom’s Hardware adds that dual channel “is vital if you’re using an iGPU”, so a desktop with no graphics card has the most to lose. One stick is playable and not half the speed, and it is still the wrong thing to buy unless the budget forces it. Ways to pay less covers it as a stopgap.

Two sticks against four

These are the speeds the processor makers will stand behind. Every stick you add loads the path to memory, and past a point the only way to keep it working is to slow it down. Intel’s overclocking guide says “many motherboards have difficulty maintaining higher speeds as the load across the memory slots is increased”.

The AMD rows are AMD’s product pages, read November 2, 2026, which print one figure per stick count with no slot term. The Intel rows are its datasheets, Volume 1: Document 655258 Revision 011 (May 2025), Document 743844 Revision 015 (May 2025) and Document 832586 Revision 007 (March 2026). Intel’s table for DDR4 gives one figure for two slots per channel with no rank split. Intel also notes that lower models may run slower, and that some 13th and 14th Gen chips follow 12th Gen’s speeds, so read your exact chip on RAM by processor. Intel lists dual-rank 5600 on 14th Gen with a note that it “is pending Intel enablement”.

• Four sticks cost AM5 a third of its rated speed. From 5600 to 3600 is 36% less, and from 5200 to 3600 is 31% less. Two modules, not four works this into capacity steps.
• On Intel 12th to 14th Gen, an empty slot costs speed too. Two sticks on a four-slot board are rated at 4400 where a two-slot board gets 4800 or 5600. Intel’s table for Core Ultra 200S shows no such penalty. The reason is in the speed guide.
• Intel’s DDR4 is the exception. Its tables print 3200 at two sticks and at four, so on that board the case for two is the free slots and the tested set.
• A rating is not a wall. Tom’s Hardware and TechSpot both tested two sticks at DDR5-7200 on an MSI Z790 board with four slots, and MSI’s manual for the MAG B650 Tomahawk WiFi lists overclocks to 6400+ with four single-rank sticks. No controlled test of 2 x 32GB against 4 x 16GB at one profile turned up in the research for this guide, so the size of that gap is unknown.

Buy four sticks only when two cannot reach the total. DDR5 has 24GB, 32GB, 48GB and 64GB sticks, so most totals fit in two slots. Four can look cheaper: on eBay US on November 1 the median new 4 x 16GB DDR5 desktop kit asked $23.55 per gigabyte against $29.13 for 2 x 32GB, both across mixed speeds. That saving buys the lower rating and leaves no slot free.

Ranks barely change this. A rank is a set of chips the controller addresses at once, and a large stick usually has two. AMD prints the same speed for both kinds at two sticks, and on Ryzen 5000 Tom’s Hardware (March 15, 2021) measured dual rank at up to 7% ahead in applications and “almost zero” in games. Single or dual rank has the rules.

Loose sticks can be cheaper, and what you give up

A kit is not always the low price. On November 1 the median new single 16GB DDR5-5600 stick on eBay US asked $21.87 per gigabyte, against $25.76 for a 2 x 16GB DDR5-5600 kit. Many of those singles are plain standard-speed parts, CL46 ones among them, and some are sold as replacements for a named prebuilt.

Two plain sticks of one part number are a fair route. They run at their standard speed with no profile, so 5600 sits inside AMD’s rating for Ryzen 9000 and Intel’s for Core Ultra 200S. You give up the tested set and some latency: a plain 5600 CL46 stick is 16.43 ns against 10 to 12 ns for a 6000 kit, which costs several percent in processor-limited games on a chip without X3D, as the tests below show.

XMP and EXPO: the speed on the box is off until you turn it on

A kit’s box speed is not what it runs at when you first start the machine. Kingston’s datasheet for its FURY Beast kit KF560C30BBEK2-32, 2 x 16GB of single-rank sticks, shows it in three lines:

Standard (JEDEC)     DDR5-4800  CL40-39-39  1.1 V
First profile        DDR5-6000  CL30-36-36  1.4 V
Second profile       DDR5-5600  CL40-40-40  1.25 V

JEDEC is the industry’s standards body, and its speeds are what any board can start at. MSI’s guide of July 21, 2026 says “During the initial boot, all motherboards will operate memory modules at ‘JEDEC’ default settings”. Corsair’s page for CMK32GX5M2B6000C36 lists an SPD speed of 4800 and latency of 40 beside a tested 6000 and 36 (the SPD is the small chip on each stick that stores these settings), and G.Skill prints “Requires overclocking/BIOS adjustments” under its rated speeds. Nothing is wrong with a 6000 kit that shows 4800. It has not been asked to do anything else.

Leaving it off costs real speed on a processor that wants memory. FPS Review (August 6, 2026) measured a Ryzen 9 9950X at the default 4800 and found DDR5-6000 CL28 ahead by 6% in Battlefield 6 and 24% in Horizon Forbidden West.

Turning it on is an overclock, and both processor makers say so. AMD’s footnote says running outside its published specifications “will void any applicable AMD product warranty, even when enabled via AMD hardware and/or software”. Intel says altering clock frequency or voltage “may void any product warranties and reduce stability, security, performance, and life of the processor and other components”. The kit is sold to run this way, and the controller being pushed is inside your processor, as the speed guide explains. To stay inside the published figures, buy a plain stick at the rated speed.

The menu names

Press Delete as the machine starts (some makers use F2) to reach the setup screen, also called the BIOS or UEFI. Then:

Menus shift with the board and the firmware version, and all three makers say so, so treat your own manual as the last word. The Gigabyte names come from its AM5 manual, and its Intel boards may word them differently; its press release of September 16, 2022 says its Z690, B660 and later Intel boards take XMP and EXPO kits.

A kit can carry either format or both, and the name on the menu depends on the pairing. ASUS publishes the table:

So an XMP kit on an AMD board is normal, and AMD’s memory compatibility list marks kits compatible with XMP. The format your platform names is still the easy choice, and some kits, such as the Kingston above, carry both.

If your AM5 board is an early one, bought used or from old stock, update its firmware before you turn EXPO on. In April 2023 AMD said a new firmware layer capped SoC voltage at 1.3 V to keep chips inside their limits, and that it did not affect overclocking memory with EXPO or XMP (quoted by PC Gamer). A board never updated does not have that fix, so get it first, and follow the maker’s flashing instructions, because Gigabyte among others warns that flashing is risky. If Windows has encrypted the drive, find your BitLocker recovery key first, because a firmware update can trigger a recovery prompt.

Boards that cannot do it

Some boards cannot be relied on to run a profile at all, so expect the kit to stay at a standard speed there. Intel’s product pages list H610 and H810 at “Supports Memory Overclocking: No” with one stick per channel, and B660 and B760 at “Yes” with two. On AMD, EXPO is general on the X and B series, and AMD says the A620, A620A and PRO 600 family supports it “on select chipsets”, so check the board. A prebuilt from a big maker may not show the setting, so ask the maker before you pay for profile speed.

Checking it took

Read the speed back after every firmware change, because a board can fall back to a standard speed without saying so, and an update can clear the setting. Three places show it: Task Manager under Performance then Memory, the board’s EZ Mode screen, and the free tool CPU-Z, whose Memory tab shows what is running and whose SPD tab shows the stored profiles. From PowerShell:

Get-CimInstance Win32_PhysicalMemory | Select-Object BankLabel, PartNumber, Capacity, Speed, ConfiguredClockSpeed, ConfiguredVoltage

ConfiguredClockSpeed is the running speed; reading the speed a machine is running explains the other numbers you may meet.

Speed by platform, and what the tests measured

Almost every figure below is processor-limited on purpose: low resolution, a very fast graphics card, the memory the only change. Those are the conditions that show the largest gaps. On an ordinary card at high resolution the gaps shrink or vanish: in the April 2025 TechSpot test below, 4K narrowed Cyberpunk 2077’s average to about 5% between the slowest and fastest kits.

AM5: the controller’s clock sets the target

Inside the processor, the memory controller has its own clock. TechSpot (April 7, 2025) explains the AM5 case: at DDR5-6000 the memory clock is 3,000 MHz and the controller can match it, a 1:1 ratio. Above that most chips cannot hold 1:1, so the controller drops to half the memory clock. DDR5-8000 therefore runs the controller at 2,000 MHz, a third slower than at DDR5-6000, in exchange for a third more bandwidth. TechSpot (November 6, 2024) calls DDR5-6200 to 6400 at 1:1 “possible but not guaranteed”.

That is why 6000 is the target, and it is not AMD’s specification. AMD told reviewers DDR5-6000 was the sweet spot, TechSpot reports, while AMD’s product pages print 5600 for Ryzen 9000 and 5200 for Ryzen 7000. AMD’s clocks has the mechanism.

TechSpot’s verdict in April 2025 was 6000 CL30, and FPS Review’s in August 2026 was the lowest-latency 6000 you can get if gaming is the goal. Neither found speed above 6000 worth a large premium in games.

A board’s “8000+” is not a promise either. TechSpot’s roundup, reported in the same April 2025 piece, found about half of 21 X870 and X870E boards ran DDR5-8000 stably, while the rest were unstable or failed to boot.

AM4: 3200, or 3600 if the board lists it

On AM4 the fabric, controller and memory clocks want to run together at 1:1:1. Tom’s Hardware (March 15, 2021) calls DDR4-3200, the official speed on Ryzen 5000, the starting point and DDR4-3600 the sweet spot, and says most Ryzen 5000 chips can hold 1:1:1 up to DDR4-4000, though AMD does not guarantee it. The gains were small: its overall figure on a Ryzen 9 5900X put 3600 at 1.3% over 3200 and 4000 at 1.8% over 3200. In games 3600 was 5.6% over DDR4-2133, with no direct figure against 3200.

DDR4-3600 CL18 and DDR4-3200 CL16 are also the same 10.00 ns. So the site’s default in adding DDR4 capacity stands: DDR4-3200, with 3600 only on a board whose memory list includes it and at a small premium.

Intel 12th to 14th Gen: gears, and a board of one kind

Intel calls the ratio a gear: Gear 1 runs the controller at the memory clock, Gear 2 at half and Gear 4 at a quarter. Its datasheet table shows DDR4-3200 in Gear 1 at the top setting and every DDR5 speed in Gear 2. This socket’s boards take DDR4 or DDR5, nearly always one or the other, so the board decides before the kit does. The socket decides covers it.

The second row is the reason for the DDR4 default. A faster DDR4 kit can run slower, because it pushes the controller into Gear 2, and TechSpot’s locked 12400F would not hold 3600 in Gear 1 at all: it locked up loading Windows. Buy 3200 on a DDR4 board.

No recent clean test of DDR5-5600 against DDR5-6000 on these chips turned up in the research for this guide. The DDR5-6000 default here rests on the tests at higher speeds, on convention and on the price rule, so do not pay much for it.

Core Ultra 200S: fast memory, little return

Intel’s datasheet (March 2026) rates plain sticks at 5600, and shows Gear 2 at the top setting and no Gear 1 at all. The 6400 on the Core Ultra 9 285K’s product page and the 7200 on the Plus chips apply to CUDIMM, a different kind of stick.

PC Gamer’s author treats anything under 5% as margin of error. On these chips the jump from 6000 to 7200 or 8200 is small, so spend the difference on capacity or on the rest of the machine.

X3D chips and work other than games

X3D chips carry extra cache, so they go to memory less often. How much less it matters is disputed. AMD’s slide for the Ryzen 7 9850X3D (reported by Club386, January 26, 2026) shows DDR5-6000 beating DDR5-4800 by 0.2% to 1.6% in five games. PC Games Hardware (January 28, 2026) tested the same question and replied that AMD showed only average frame rates: the lows moved by up to nearly 20% and single averages by about 14%. TechSpot’s July 2026 test also found 9% and 11% on a 9800X3D in the two most processor-bound of five games, from timings alone.

Work other than games has its own pattern. Phoronix (January 29, 2026), on the same chip under Linux, found 300-plus benchmarks only just over 1% apart by geometric mean, but 5 to 10% better on DDR5-6000 in code compilation, technical computing, database serving and AV1 encoding. FPS Review found DDR5-8000 best at video transcoding, with Cinebench and Blender about level. If the machine compiles or encodes for a living, data rate matters more; what CAS latency is worth sorts workloads the same way.

Compare latency in nanoseconds, not CL

CL counts clock cycles, and a faster kit has shorter cycles, so a bigger CL is not always slower. The latency guide owns this subject, and the working takes one line:

latency in ns = CL x 2000 / data rate

The CL on a box is the rated figure with the profile on. TechSpot (February 18, 2026) found a CXMT-based DDR5-6000 CL36 kit matched a premium CL26 kit in games. The price gap can be wide: on November 1, among new 2 x 16GB DDR5-6000 kits with a CL this site can read, CL30 had a median of $27.22 per gigabyte and CL36 $15.06. That rests on 24 and 25 listings, and only about one kit in eight gave a readable CL, so treat it as a hint. If the CL30 premium would buy the next capacity step, the capacity is the better purchase, as the CL30 and CL36 comparison argues.

CL is also not the whole story. AMD’s new EXPO 1.2 standard adds “Ultra Low Latency” profiles that tighten the timings no box prints. TechSpot’s test of July 15, 2026 used G.Skill kits on a Ryzen 7 9700X at 1080p. Over a standard CL30 kit, the CL30 version gained 8% to 15%, and the CL36 version 7% to 9% in the four games where TechSpot states it, so a CL36 kit beat a CL30 kit on timings the box does not show. On a 9800X3D the gain was mostly 1 to 3%, and at 4K mostly nothing. AMD claims 4%. The tested CL36 kit cost about 6% more than a comparable standard kit, though early reports ran as high as 80%, and support needs the right firmware: Tom’s Hardware (April 2026) reported EXPO 1.2 on 800-series boards first, so check your board’s firmware notes.

CUDIMM, and two cheap kinds of stick to read for

CUDIMM. A CUDIMM carries a small clock driver chip that cleans up the signal at high speed. It fits the same slot and is still an unbuffered desktop stick. The clock driver explains it.

• On AM5 it is not worth paying for. Ryzen 7000, 8000G and 9000 chips run it only in bypass mode, limited to around 6000 MT/s, Tom’s Hardware reported on April 28, 2026. Native support is “reportedly planned” for Zen 6.
• On Core Ultra 200S it works natively, within limits. Intel rates CUDIMMs at 6400, and at 7200 on the 200S Plus, in the column its note defines as boards that route one slot per channel. In its column for boards with two slots per channel, the same table prints 5600 for plain sticks and CUDIMMs alike. Read literally, a four-slot board’s rated speed does not rise, and the 6400 becomes a board-maker figure, so check the board’s memory list.
• On Intel 12th to 14th Gen it is not listed. Those datasheets name no CUDIMM, so do not pay for one on an LGA1700 board.
• The gain is small either way. TechSpot (October 26, 2025): “High-speed CUDIMM isn’t a silver bullet.” Its March 2026 test put an 8200 CUDIMM kit 1% to 5% over 6000 in three games.

HUDIMM. This is a new kind of stick with half the usual chips: one 32-bit subchannel where normal DDR5 has two, so half the capacity and half the bandwidth. Tom’s Hardware (April 19 and 20, 2026) reported ASRock and TeamGroup behind it, on Intel 600, 700 and 800 series boards, and relayed a simulation with taped sticks that read 58 GB/s against 106 GB/s in dual channel. It had no price or availability then, and whether any has reached shops was not confirmed here. A stick much cheaper than its neighbors at the same size deserves a look at its specification.

A “1+1” pack. A V-Color pack of that name is one working stick and one lit dummy (Tom’s Hardware, September 4, 2026), and the guide to paying less lists it among the offers that only look cheap. It gives you one channel.

How much to buy

How much RAM do you need owns this question, by workload. For a kit buyer, the summary is short: 32GB as two sticks for a new gaming or general build, 64GB as 2 x 32GB if you already know you will need it, and 16GB for an office or esports machine, or if the budget leaves no choice.

Those are the DDR5 desktop page’s medians with New and each size chosen, before shipping.

A 2 x 8GB kit is a dead end. A later upgrade means four sticks, which on AM5 means a rated 3600, or selling the pair and paying twice. At those medians the second 16GB costs about $25 a gigabyte ($883 less $480, over 16GB), less than the average gigabyte of either kit. TechSpot (July 15, 2026) would buy 2 x 8GB to get through the shortage, and 16GB does still run most games, but that does not change what happens at the next upgrade. If 32GB is out of reach, the capacity guide’s routes price the alternatives.

DDR5’s 24GB and 48GB sticks reach 48GB and 96GB in two slots at the full two-stick speed; the 24GB and 48GB pages list them, laptop sticks among them. Read the manual’s largest stick first: Gigabyte’s for the B650 Gaming X AX stops at 48GB, 192GB in all. On DDR4, unbuffered desktop sticks stop at 32GB, per the DDR4 guide.

Adding to what you already have

Replacing a kit is the clean answer, and adding one is a gamble you can sometimes win.

This does not mean mixed sticks refuse to run. The mixing table is right that nothing in memory training reads a part number, and that “you must use identical part numbers” is folklore. Intel’s note makes a narrower point: the machine can start, and the top speed is no longer promised. Expect mixed sticks to run at a standard speed, and treat any profile as a bonus.

If you do add, the order that loses least comes from Tom’s Hardware. Test at the standard speed first with the profile off, and run your usual programs and a memory test for a few days. Then turn the profile on and repeat, inside the return window. Fit four sticks only with one type and the same rank count, since channels with mismatched ranks lose out.

Or replace the pair. Two larger sticks keep the two-stick rating and leave two slots free, and the old kit has resale value: read the used rows near the median and sell it the month you buy. On DDR4, the paths from 2 x 8GB are priced.

One board feature changes the sum. ASUS (Edge Up, April 27, 2026) says AEMP II and III on its Z890 and B860 boards, with firmware 3002 or 3103, can tune mixed plain sticks: its demonstration took four sticks across three brands to 5200 on a Core Ultra 7 265K. It works only on sticks with no XMP or EXPO profile, all of one rank type.

Checking your board before you pay

This site has no page for a desktop board, and no guide here can tell you that a given kit fits one. Four things in the maker’s documents do most of the work.

1. The memory type. DDR4 and DDR5 are notched differently and one will not enter the other’s slot. AM4 is DDR4, and AM5 and LGA1851 are DDR5. Read the rest in form factor and keying.

2. Which slots to fill. Almost all four-slot boards want the second slot on each channel. MSI: “Always insert memory modules in the DIMMA2 slot first”, and two go in DIMMA2 and DIMMB2. ASUS: one in DIMM_A2, two in DIMM_A2 and DIMM_B2, and search the manual for “Recommended memory configuration”. Gigabyte: DDR5_A2 first, two in DDR5_A2 and DDR5_B2. Intel’s note gives the same rule: “Far memory slot to be populated”. Slots A1 and B1 alone can train lower or not at all. A board may differ, so read yours.

3. What the board’s speed figure means. A figure on a box or in a manual is usually the board’s overclocking ceiling, in one best-case layout. Three entries, each read from the maker’s own manual:

4. The memory list. Each board maker keeps a memory support list, also called a QVL, naming kits it tested at a speed and a stick count. QVL lists, and what they are worth has the fine print: it shows what has been seen working, and absence from it is not a verdict. Prefer a kit on it if the price is the same. Three other checkers answer from a different side, none of them this site: AMD’s Ryzen Processor Overclocked Memory Compatibility List (“memory kits tested across systems to run at rated speed and latency”), G.Skill’s RAM Configurator, and Corsair’s RAM and Motherboard Compatibility Checker.

Height, heat, lights and software

Height. A plain stick stands about 30 mm. Kingston’s FURY Beast DDR5-6000 is 34.9 mm with its heatsink, and tall finned kits stand higher. A large tower cooler can overhang the slots, and its maker publishes a RAM clearance figure for that model. Compare that figure with the height on the kit’s own datasheet before you buy either. A slim desktop has the same problem at its lid.

Heat. TechSpot (February 18, 2026) saw two kits peak at 55 and 60 degrees after an hour of Prime95, on case airflow alone. In its experience higher-speed DDR5 “can run into stability issues above 70”, against a JEDEC limit of 95, though it says this is not normally something it would worry about. ASUS’s manual asks for better memory cooling under a full memory load or overclocking.

Lights and software. RGB adds nothing to speed. G.Skill says its kits are controlled through its own software or through Asus Aura Sync, Gigabyte RGB Fusion, MSI Mystic Light or ASRock Polychrome Sync, and Corsair’s kits use iCUE. The profile is a firmware setting, and the makers’ instructions for it ask for none of this software.

ECC and registered sticks: the classic used-market mistake

A registered stick (RDIMM) does not run in an ordinary desktop board, and this is the classic mistake with used memory. Sticks pulled from retired servers are plentiful and can look cheap, and on DDR4 a registered stick has the same 288 pins and the same notch as a desktop one. It goes in, and the result is fans spinning and a dark screen, because a desktop’s firmware has no code to set up the register. Why the wrong type is a dead board walks through it. On DDR5 the two types have different notches, so the wrong one will not seat. AMD’s desktop processor pages list “UDIMM” as the memory subtype, meaning unbuffered, and the manuals above say un-buffered in as many words.

The price is the lure, and the saving is small while the mistake is complete. On a DDR4 label look for R straight after the hyphen that follows the speed, as in PC4-2666V-RB2; a letter joined to the speed, as in PC4-2400R-UA1, is only a latency bin, and that stick is unbuffered. The used-RAM guide’s classic mistake has the prices and the part-number tells. The registered DDR4 and registered DDR5 pages exist for servers, and the desktop tables here leave out sticks whose listing states ECC.

ECC is a separate question. Unbuffered ECC is a real desktop option on some boards and not others. AMD’s pages say “Yes (Requires mobo support)”, and the AM5 manuals above disagree, which is why you read the board and not the socket. On Intel, a chip’s product page can say yes while the consumer chipset says no. Which CPUs and boards support ECC has the table, and who needs ECC will tell you if you are one of them.

First boot: long training, and what to do when it will not start

Fitting. Turn the machine off, unplug it from the wall and touch a metal part of the case to discharge static, as ASUS, MSI and Gigabyte all say to, and hold each stick by its edges. Open the slot’s clip, line the notch up, and push straight down until the clip locks: Corsair says you should hear a click. A stick goes in only one way, so if it will not go, turn it around instead of pressing harder.

Training. On the first start the board measures and adjusts its connection to each stick, and on DDR5 that can take minutes. Gigabyte’s firmware notes say “Higher memory speeds may cause longer memory training time on the first boot”. A sticker on an ASRock X670E board, reported by Guru3D in September 2022, gave these figures from launch-era firmware, which ASRock said it was shortening:

MSI’s advice for fans spinning and no picture, “especially for AMD Ryzen systems”, is “to just wait for a while”. So give a first start several minutes before you touch the power button. Training runs again after you clear the settings, update firmware, change the speed or move a stick, as boot-time training explains. If every start stays long, look for Memory Context Restore: MSI and Gigabyte both describe it as cutting that wait.

When it still will not start, work down this table.

If one stick works alone and the other does not in any slot, you have found a dead stick, so use the return window. If both work alone and not together at the profile, the kit may be fine at a lower speed.

Testing the kit

The full order is in the week it arrives, which serves new kits too. In short: confirm the total and speed in Task Manager, compare the part number with the label, then run MemTest86 from a USB stick at the profile you mean to use. The two testers compares it with Memtest86+.

Errors with the profile on and none at the standard speed usually point at the profile or the processor’s memory controller, and not at a dead stick. Run the kit slower or use the seller’s return policy. Do it in the first week, because the return clock starts at delivery.

New or used

A used 2 x 16GB kit asked about half the new price on DDR5 and about two-thirds on DDR4. Used DDR5 desktop kits are scarce, though, so a good price may not be there on the day you look.

Those are the medians the same desktop pages printed with Used or New chosen: item prices, before shipping. A used kit is worth buying when both sticks came out of one kit, the part number can be read, and you can test it inside a return window. The sums on that last condition are unforgiving: at the DDR5 median, a used kit with one dead stick you cannot return costs $28.12 for each working gigabyte, a little over the new price.

Four things are specific to a desktop kit:

• The heat spreader hides the chips, so the part number on the label does all the work. Packaging and heat spreaders has the rest.
• The headline speed is a profile, so a seller’s “tested working” may mean at the standard speed. Test at the speed you mean to use.
• “Overclocked” is normal use, because every enthusiast kit is rated as an overclock.
• A maker’s warranty may not follow the kit, as the warranty section of the DDR4 guide sets out.

A pair of sticks from two sellers is two loose sticks, whatever the title calls it. Buying used RAM on eBay covers the rest, including what a used desktop listing looks like and a worked desktop kit, and compare Buy It Now prices, because a bid is not a price.

Five desktops, read from the makers’ documents

These are read from the makers’ own pages to show the method, not a claim that this site knows what fits a board, and a kit named here is no best buy. Check any kit against your own board’s list.

A new AM5 gaming build. AMD’s page for the Ryzen 7 9800X3D gives UDIMM, 256GB, 5600 with two sticks of either rank and 3600 with four. MSI’s manual for the MAG B650 Tomahawk WiFi gives four slots, non-ECC unbuffered memory only, and DIMMA2 and DIMMB2 for two sticks. Kingston’s KF560C30BBEK2-32 datasheet gives 2 x 16GB, single rank, 4800 at first and an EXPO profile of 6000 CL30 at 1.4 V, at 34.9 mm tall. Read together: two sticks, 6000 is an overclock above AMD’s 5600, and the EXPO profile is chosen at the top of EZ Mode. On an X3D chip, CL36 is the place to save if CL30 costs noticeably more.

A Ryzen 5 7600X on a Gigabyte B650 Gaming X AX. AMD’s page says 5200 with two sticks and 3600 with four, 128GB. Gigabyte’s manual prints 5200, 4800 and 4400 and supports EXPO and XMP, 48GB a stick, DDR5_A2 and DDR5_B2. This chip has no X3D cache and lost the most in Tom’s Hardware’s one-stick test, so memory matters to it. Buy 2 x 16GB DDR5-6000 if it is within a few percent of 5600, then use Tweaker, XMP/EXPO Profile.

A Core i9-14900K in a four-slot board. Intel’s datasheet gives DDR5 at 5600, or 4400 on a two-slot-per-channel board, 4000 or 3600 with four sticks, and DDR4 at 3200, and the chip’s page carries the summary. The board decides between the two generations. On a DDR5 board, Corsair’s CMK32GX5M2B6000C36 is a plain example of an XMP kit: 4800 CL40 at first, tested at 6000 CL36, listed for Intel 600, 700 and 800 series. On a DDR4 board buy 3200 CL16, keep the board, and read the drop-in sum before you consider changing it.

A Core Ultra 9 285K in a four-slot board. Intel’s table gives 5600 with one plain stick per channel, 4800 or 4400 with four sticks, and Gear 2 at best. PC Gamer’s five games put 7200 within 2% of 6000 on this chip. So 2 x 16GB DDR5-6000 with XMP is enough, and a CUDIMM kit is not needed unless it costs about the same.

A Ryzen 7 5800X3D owner with 2 x 8GB. AMD’s page for this chip gives DDR4-3200 with two sticks, 2933 with four single-rank and 2667 with four dual-rank, and 128GB. Adding a second 2 x 8GB gives 32GB at a lower rated speed, from two kits nobody tested together. Replacing the pair with 2 x 16GB keeps 3200 and frees two slots. On November 1 the median used 2 x 16GB DDR4-3200 kit asked $5.94 per gigabyte, about $190, and new $10.76, about $344.

What cannot be known before you buy

• Whether your chip holds the profile. The controller is in your particular processor, and two of the same model can differ. TechSpot’s “possible but not guaranteed” at 6200 to 6400 on AM5 is this.
• Whether a kit trains on your board at your stick count. A list shows that someone saw it work on some firmware. That is good evidence, not a guarantee.
• What a used kit has been through. No odometer exists, which is why the test comes first.

So buy from a seller with a return window, and test in the first week. A wrong kit is cheap only if you can send it back.