In this guide: 21 sections
- Who should still buy DDR4, and who should not
- Before you pay for DDR4: the checklist
- DDR4 more than doubled in a year, and it is still the cheaper gigabyte
- The chip now costs more than DDR5’s, and the kit has not caught up yet
- Nobody is building the DDR4 you would want to buy
- Both DDR4 platforms got a second life, and higher prices with it
- On the same processor, DDR4 gives up 9% to 17% in CPU-bound games
- Dollars per percentage point: the sum that settles it
- If you already own the DDR4 machine, the drop-in beats the rebuild
- Adding DDR4 capacity: fill the slots once, and read what four modules cost
- Windows 10 ended in October 2025, and the oldest DDR4 machines are on borrowed time
- DDR4 laptops: about half the retail price per gigabyte and a 64GB ceiling
- On the trackers, server DDR5 costs four to six times server DDR4 per gigabyte
- A DDR4 RDIMM fits a desktop slot and will not start the machine
- Used memory comes from five places, and the server stream is starting to keep its own
- Ex-server modules have the best history on the market, and one gap in it
- “Ex-mining” means three workloads, and only one of them rewarded overclocking the memory
- Every enthusiast kit is rated as an overclock, and no module remembers how it was run
- “New” is not a provenance either
- The warranty stays with the first owner, and G.Skill’s “lifetime” ends with the model
- Filter by the machine first: which page on this site answers which question
Yes for a machine you already own: DDR4 is still the cheapest mainstream memory, at roughly half DDR5’s US retail price per gigabyte or less. For a new build it is the budget route, with clear terms: DDR4 kits cost 2.2 to 2.8 times what they did a year earlier, its chips now cost more than DDR5’s, and on the one processor family that takes both generations DDR5 leads by about 10% to 20% in CPU-bound games. It pays best when the graphics card is the limit or the board is already owned, and least on a new high-end processor. Used DDR4 is a sound buy when its origin is read rather than assumed and it is tested inside the return window, which a “no returns” listing on eBay still has.
Prices below date from March to late September 2026, each with its source and date; where a figure is worked out here, the arithmetic is shown so you can change the inputs. Why the shortage happened is in the 2026 DRAM shortage, and when to buy is in buy RAM now or wait.
Who should still buy DDR4, and who should not
| Your situation | Verdict | The deciding figure | Details |
|---|---|---|---|
| You own an AM4 or LGA1700 DDR4 machine and want it faster | Keep the DDR4 and drop in a processor | A $349 Ryzen 7 5800X3D; switching to AM5 at 16GB costs $120 more than that, plus the board, less resale | The drop-in sum |
| You own a DDR4 machine and want more capacity | Buy DDR4, once, now | DDR4 desktop modules stop at 32GB, four slots at 128GB; four modules cost 8% to 17% of peak bandwidth | Adding capacity |
| You want to sell your DDR4 to fund DDR5 | Usually not worth it on its own | Even at the $190 of the best new 32GB kit, 32GB of DDR5 at $409 costs $219 more | Selling DDR4 |
| New build, older or mid-range processor, or the graphics card is the limit | DDR4 on LGA1700 or AM4 still pays | 32GB of DDR4-3200 at $190 against $409 for DDR5-6000; each DDR5 point costs at least $23.05 on a Core i5-12600K | Dollars per point |
| New build around a high-end processor | DDR5, or a weaker processor on DDR5 | DDR4 cost a Core i9-14900K 14.0% in CPU-bound games; a DDR5 point costs at least $15.64 there | The gaming gap |
| A cheap Skylake or Kaby Lake office PC, bought for its DDR4 | Only as a stopgap | Not on Microsoft’s Windows 11 lists; consumer Extended Security Updates end October 12, 2027 | Windows 10 |
| A homelab or server | Ex-server DDR4 RDIMMs | Two trackers put server DDR5 at 4.4 to 5.6 times DDR4 per gigabyte; a DDR4 EPYC socket has a third of the bandwidth | Servers |
| A DDR4 laptop | Fill the empty second slot first | Two slots stop at 64GB; one module runs single-channel | DDR4 laptops |
| A used branded kit against a cheap new no-name one | The used branded kit | Its part number can be checked against its label, its chip packages and its SPD | “New” is not a provenance |
The pattern under the table: the cheaper DDR4 gigabyte is worth most where the memory is not the bottleneck, and least where a fast processor would be held back by it.
Before you pay for DDR4: the checklist
- Identify the machine and its Windows 11 status first. The processor, the board and Microsoft’s list decide what the memory is worth: use the CPU index, how to check what RAM fits and the Windows 10 section.
- Decide keep, add or switch with the per-point sum, using your own board quote and resale figure rather than anyone’s verdict.
- Compare at a fixed capacity using medians, rather than the cheapest rows, and on the all-generations table when crossing generations.
- Read the word “registered”, not the photograph. A DDR4 server module seats in a desktop slot and the machine does not start.
- Read the part number and the rank before the price, and match a second kit by part number on the part index.
- Fill the slots once. Two large modules keep full speed at a premium per gigabyte; four smaller ones cost less and run slower.
- For a server, look up the machine and buy by its channel count, on the server index, not by the price of the cheapest lot.
- Ask which kind of mining, if a seller mentions it, and test accordingly.
- Treat a cheap “new” no-name DDR4 kit as a claim to test; a used branded kit is often the more checkable buy.
- Price the warranty gap with the break-even sum, not a percentage rule.
- Test inside the return window. A “no returns” listing still has one: eBay’s Money Back Guarantee covers a module that is not the one described, or arrives broken, damaged or defective without being described that way, for 30 days from the actual or estimated delivery date, or the seller’s own return window if that is longer. “No returns” removes only change-of-mind returns. “Untested” is a reason to pay less, not to skip the test; the method and the return steps are in buying used RAM on eBay.
DDR4 more than doubled in a year, and it is still the cheaper gigabyte
Tom’s Hardware’s US price index, last updated on September 14, 2026, concludes that “DDR4 isn’t exactly a safe haven for consumers.” The cleanest year-on-year series is the PCPartPicker averages Tom’s Hardware compiled on August 17, 2026, which the author calls “somewhat approximate” but which cover the same kits a year apart. Per-gigabyte figures and changes are computed here.
| Kit | Aug 2025 average | Aug 2026 average | Change | Aug 2026 per GB |
|---|---|---|---|---|
| DDR4-3200 2x8GB | $63 | $163 | +159% | $10.19 |
| DDR4-3600 2x8GB | $75 | $165 | +120% | $10.31 |
| DDR4-3200 2x16GB | $105 | $281 | +168% | $8.78 |
| DDR4-3600 2x16GB | $120 | $307 | +156% | $9.59 |
| DDR4-3200 2x32GB | $222 | $614 | +177% | $9.59 |
| DDR4-3600 2x32GB | $300 | $789 | +163% | $12.33 |
| DDR5-4800 2x16GB | $90 | $425 | +372% | $13.28 |
| DDR5-5600 2x16GB | $116 | $528 | +355% | $16.50 |
| DDR5-6000 2x16GB | $108 | $572 | +430% | $17.88 |
| DDR5-5600 2x32GB | $191 | $1,118 | +485% | $17.47 |
| DDR5-6000 2x32GB | $222 | $1,272 | +473% | $19.88 |
Every DDR4 kit in the comparison rose by between 120% and 177%. DDR5 rose by more, 355% to 485%, so the ratio moved in DDR4’s favor:
2x16GB, Aug 2025 DDR4-3200 105 / 32 = $3.28/GB DDR5-6000 108 / 32 = $3.38/GB ratio 0.97
2x16GB, Aug 2026 DDR4-3200 281 / 32 = $8.78/GB DDR5-6000 572 / 32 = $17.88/GB ratio 0.49
best 32GB kit, DDR4-3200 190 / 32 = $5.94/GB DDR5-6000 409 / 32 = $12.78/GB ratio 0.46
Tom's index, Sep 14, 2026
A year ago a gigabyte of DDR4 and a gigabyte of DDR5 cost the same at US retail. By September 2026 the DDR4 gigabyte cost about half. Other dated sources agree. XDA Developers priced 32GB of DDR4-3200 at $179 against $402 for DDR5-6000 on August 26 (0.45); the Capital and Compute tracker’s September 26 snapshot put 32GB DDR4 desktop kits at $129 to $233 against $429 to $560 for DDR5 (0.30 low against low, 0.42 high against high); and eRacks, a US server builder, put the median DDR4 desktop gigabyte at $6.22 against $14.58 for DDR5 on September 22 (0.43). Five sources, five methods, and a DDR4 gigabyte at between about a third and a half of a DDR5 one.
To read the live ratio, divide the medians printed at the top of 32GB DDR4 kits and 32GB DDR5 kits, rather than the two cheapest rows. The all-generations table sorts both generations on one scale, and Trends draws how each generation’s like-for-like price per gigabyte has moved since October 3, 2026.
The chip now costs more than DDR5’s, and the kit has not caught up yet
As silicon, DDR4 now costs more per bit than DDR5. As a retail kit, it is still about half the price per gigabyte. The gap between the two decides how long the budget route lasts.
DRAMeXchange, TrendForce’s price board, published a session average of $83.784 for a DDR4 16Gb 3200 chip on September 24, 2026, against $57.667 for a DDR5 16Gb chip. A 16-gigabit chip holds 2GB, so per gigabyte of silicon:
DDR4 16Gb 3200 83.784 / 2 = $41.89 per GB of silicon
DDR5 16Gb 4800/5600 57.667 / 2 = $28.83 per GB of silicon
DDR4 as a share of DDR5 41.89 / 28.83 = 1.45
The same DDR4 line traded between $45.50 and $119.00 that day, a spread that points to a thin market, but its series is steep: DDR4-3200 8Gb chips cost $1.75 at the end of April 2025, per ComputerBase data relayed by Tom’s Hardware, and $45.54 on September 15, 2026 by TrendForce’s weekly figure, 26 times as much. The same gigabyte at each layer of the market:
| Layer (source, date) | DDR4 per GB | DDR5 per GB | DDR4 as share of DDR5 |
|---|---|---|---|
| Chip spot, 16Gb chip (DRAMeXchange, Sep 24, 2026) | $41.89 | $28.83 | 1.45 |
| Module spot, 16GB UDIMM (DRAMeXchange, Sep 14, 2026) | $10.27 | $14.73 | 0.70 |
| Retail, average 2x16GB kit (PCPartPicker via Tom’s Hardware, Aug 2026) | $8.78 | $17.88 | 0.49 |
| Retail, best 32GB kit (Tom’s Hardware index, Sep 14, 2026) | $5.94 | $12.78 | 0.46 |
On September 14 the best US retail DDR4 kit sold at 5.94 / 10.27 = 58% of the module trade price per gigabyte, while the best DDR5 kit sold at 12.78 / 14.73 = 87% of its own. Retail DDR4 is sitting much further below replacement cost than retail DDR5 is.
Two explanations fit. One is inventory: TrendForce’s September 22 report credited part of ADATA’s 2025 profit growth to “low-cost inventory” bought in the previous cycle. The other is that retail reprices slowly, for the reasons set out in the 2026 DRAM shortage. The practical reading is a mechanism, not a forecast. When the cheap inventory runs out, retail DDR4 reprices toward what a new module costs to make, and on the September figures that is well above where retail DDR4 sits. Nobody can say when. “DDR4 is the cheap option” is a statement about stock on shelves, not about the technology.
Nobody is building the DDR4 you would want to buy
Less DDR4 is being made, for other customers, on old processes.
The largest makers are pulling back, though not uniformly. Micron had sent end-of-life notices for DDR4 to its high-volume customers by mid-2025, chief executive Sanjay Mehrotra said on its June 25, 2025 earnings call, and the DDR4 it still makes at its Manassas, Virginia fab is, in its June 2026 earnings deck, for “the legacy product needs of our customers in auto, industrial, medical, aerospace and defense markets.” TrendForce described Samsung on January 19, 2026 as “sticking firmly to its end-of-life plan”, though other reports had Samsung continuing DDR4 through 2026 and delaying end of life for a server customer, so its exit is contested. SK hynix, first reported in July 2025 to be stopping DDR4 orders in October 2025, was reported in September to be boosting DDR4 at its Wuxi fabs, and moved that line to the 1a node. China’s CXMT, not one of the big three, “dropped out of DDR4 to make more DDR5”, Network World reported on June 1, 2026. TrendForce said on June 22, 2026 that the big three had cut wafer allocations for DDR4, pushing buyers to the Taiwanese makers.
Those makers are growing, so “no new consumer DDR4 is being manufactured”, a line still quoted in 2026, is wrong: Nanya is putting more capacity into DDR4 and LPDDR4 than into DDR5, and Winbond’s DDR4 and LPDDR4 shipments lifted its first-quarter DRAM revenue 91.4% quarter over quarter, TrendForce reported on June 1, 2026. But their output is sold before it reaches you. Nanya’s president said in July that customers cannot get enough DDR4 or LPDDR4 from anyone, and Winbond’s new 16nm DDR4 goes to television, networking, industrial and embedded buyers whose orders fill its capacity through 2027.
No leading-edge process will make DDR4 cheaper. Tom’s Hardware reported in August 2025 that no DDR4 would be made on the industry’s newest 10nm-class nodes, and a generation gets cheaper when its chips move to the leading edge in volume.
The best DDR4 is no longer made at all. At Computex in June 2026 more than half a dozen module and board makers told Tom’s Hardware they were shifting production back toward DDR4, with a catch: “high-performance DDR4 dies, such as the famous Samsung B-dies, are no longer in production. As such, most of the revamped DDR4 kits will top out at a rather pedestrian DDR4-3600.” Fast, tight DDR4 kits are now a finite second-hand pool, and the squeeze reaches the generation below: some manufacturers are redesigning DDR4 products to use DDR3.
Both DDR4 platforms got a second life, and higher prices with it
Two sockets make up the whole new-build budget route: AMD’s AM4, which takes only DDR4, and Intel’s LGA1700, whose 12th to 14th-generation Core processors take either generation while the board takes one. Both were written off in early 2026 (Tom’s Hardware in January, BGR in April); by September AM4 had new processors and boards on sale, and Intel had said it would keep LGA1700 in its lineup for years.
On AM4, AMD put a Ryzen 7 5800X3D “10th Anniversary Edition” on sale from June 25, 2026 at a suggested $349, $100 below the original launch price, and told Tom’s Hardware it would keep selling it “as long as it makes sense”. A $99 Ryzen 5 5500F followed in September, according to TechSpot, and Gigabyte launched two new B450 boards in August with four DDR4 slots and support for up to 128GB at DDR4-2667. PCIe, the other platform limit people raise, is the smaller one: XDA Developers’ August 26, 2026 case for AM4 cited GamersNexus’s RTX 5090 tests, in which PCIe 5.0 against 3.0 “amounted to a minuscule 1-4%”. On LGA1700, Intel’s Robert Hallock called Raptor Lake in August “a core part of the portfolio that I want to offer to people for years to come”, and at least two board makers said they were ramping DDR4 boards into 2027.
| Platform | Example CPU (price, date) | Official DDR4 speed | Maximum memory | Status and source |
|---|---|---|---|---|
| AM4 | Ryzen 7 5800X3D ($349 suggested, from Jun 25, 2026) | 3200 with 2 modules; 2933 with 4 single-rank; 2667 with 4 dual-rank | 128GB | Re-released, not a limited run (AMD via Tom’s Hardware, Jun 4, 2026) |
| AM4 | Ryzen 5 5500 ($80, May 2026); Ryzen 5 5500F ($99, September 2026) | Read the CPU and board pages | - | New Gigabyte B450 boards, 128GB at DDR4-2667 (Tom’s Hardware, Aug 14, 2026) |
| LGA1700, DDR4 board | Core i5-14600K (about $250, August 2026) | DDR4-3200 or DDR5-5600, decided by the board | 192GB | “Core part of the portfolio” (Intel’s Hallock, Aug 14, 2026); “Raptor Lake Next” reported for H1 2027, Intel declined to comment |
| LGA1700, budget | Core i3-12100F ($90), Core i3-14100F ($100), May 2026 | DDR4-3200 or DDR5-4800, decided by the board | 128GB (12100F), 192GB (14100F) | Two board makers ramping DDR4 boards into 2027 (Tom’s Hardware, Jun 4, 2026) |
The AM4 speeds are AMD’s for the 5800X3D only, and the LGA1700 rows are Intel’s ARK entries, both read on September 26, 2026. On LGA1700 the board, not the processor, fixes the generation for good, which DDR4 vs DDR5 explains.
The revival had a price. Tom’s Hardware reported in August that the i5-14600K, “$200 or less for the better part of last year”, had “jumped to around $250, if you can find it in stock at all”, and found the i7-14700K at $380 against about $330 expected. The platforms are open, but no longer the cheap part of the build.
On the same processor, DDR4 gives up 9% to 17% in CPU-bound games
Tom’s Hardware retested eight LGA1700 processors in July 2026 with the memory as the only variable: an RTX 5090, 1080p, fifteen games, XMP enabled on both. The loss grows with the processor:
| CPU | DDR4-3200 (4x8GB) against DDR5-7200 (2x16GB), 1080p geomean |
|---|---|
| Core i9-14900K | -14.0% |
| Core i7-13700K | -13.5% |
| Core i7-14700K | -13.2% |
| Core i5-14600K | -13.1% |
| Core i5-13600K | -11.4% |
| Core i9-13900K | -11.3% |
| Core i5-12600K | -9.5% |
| Core i7-12700K | -9.3% |
Single games went further: Marvel Rivals lost 25.3% on the 14700K. And DDR4 flattens a processor upgrade: on DDR4 the 14900K was only 1.7% faster than the 13900K, against 4.6% on DDR5. Read the bench before the headline. It paired DDR5-7200, a fast kit, with mainstream DDR4-3200 in four 8GB modules rather than two, so it measures that pairing, not every DDR4 build.
TechSpot’s December 2025 retest used speeds closer to what people buy, though its DDR4 was a premium CL14 kit of the kind no longer made: DDR4-3600 against DDR5-6000 on the same Core i5-12400F, twelve games, again an RTX 5090 at 1080p. It found DDR5 “typically delivers a ~20% performance bump over DDR4 (on average)”, which as DDR4’s loss is 1 - 1/1.20 = about 17%. The games and the processor differ between the two tests, hence the range.
The most awkward result for DDR4 came three days after Tom’s Hardware’s July gap test. The same outlet’s single-DIMM test found that one 16GB DDR5 module, which loses 8-11% against two on most processors and only 2.8% on a Ryzen 7 9800X3D, “offers superior performance to dual-channel DDR4 on CPUs that support both memory generations”: on the 14600K a single DDR5-7200 module beat dual-channel DDR4-3200 by 6.8% in 007 First Light, with half the capacity of the dual-channel runs, though “With slower speeds, the performance will equalize”.
Every test above is CPU-bound by design. The nearest evidence beyond that is a whole-system comparison: Club386 (February 6, 2026) ran a Ryzen 7 5700X with DDR4-3200 against a Ryzen 7 9700X with DDR5-6000, both with 32GB and the same RTX 5070. Where the processor was the limit the DDR5 machine was far ahead, by 132-135% in Rainbow Six Siege X at medium settings; where the graphics card was the limit the gap nearly vanished, 1-3% in Assassin’s Creed Shadows and 3-4% in Cyberpunk 2077 with ray tracing. The processor changed too, but the shape is clear: once the graphics card is the limit, the memory generation stops mattering much. Why games are the sensitive case is in the desktop kit guide.
Dollars per percentage point: the sum that settles it
Tom’s Hardware’s July verdict held both sides: “Dollars spent for performance gained, DDR4 makes complete sense”, and “It’s a dead end, both for performance and upgrade potential.” To decide between them, put the price and the penalty in one number. On a new LGA1700 build, with the memory as the only difference:
Tom's Hardware pairing, Core i5-14600K
memory difference 32GB DDR5-6000 409 - 32GB DDR4-3200 190 = $219 (index, Sep 14, 2026)
DDR4 penalty = 13.1% (test, Jul 28, 2026)
per point 219 / 13.1 = $16.7
TechSpot pairing, Core i5-12400F
memory difference DDR5-6000 2x16GB 572 - DDR4-3600 307 = $265 (PCPartPicker, Aug 2026)
DDR4 penalty DDR5 about 20% faster: 1 - 1/1.20 = about 16.7%
per point 265 / 16.7 = $15.9
The 13.1% was measured against DDR5-7200, while the $219 prices DDR5-6000, so $16.7 a point, and every figure in the table below, is a floor on what a DDR5 point costs. Across the eight processors in the July test, the same $219 says more than either verdict: the weaker the processor, the more each point of DDR5 costs, because a weaker processor loses less to DDR4.
| CPU | DDR4 penalty | Dollars per point at a $219 memory difference |
|---|---|---|
| Core i9-14900K | 14.0% | $15.64 |
| Core i7-13700K | 13.5% | $16.22 |
| Core i7-14700K | 13.2% | $16.59 |
| Core i5-14600K | 13.1% | $16.72 |
| Core i5-13600K | 11.4% | $19.21 |
| Core i9-13900K | 11.3% | $19.38 |
| Core i5-12600K | 9.5% | $23.05 |
| Core i7-12700K | 9.3% | $23.55 |
That is the memory alone. A DDR5 board and a DDR4 board for the same socket do not cost the same, so the board enters as your own quote:
with boards (219 + B5 - B4) / 13.1 B5, B4 = your DDR5 and DDR4 board quotes
every $50 of board difference adds 50 / 13.1 = $3.8 per point
There is a third route, Tom’s Hardware’s own advice: “bite the bullet on DDR5 and buy a weaker CPU.” A 16GB DDR5-6000 kit was $239 on the September 14 index, $49 more than the $190 32GB DDR4 kit, for half the capacity. The index does not say whether that kit is one module or two (the usual 16GB kit is 2x8GB), so read the module count on the listing. Whether 16GB is enough is answered in how much RAM you need.
| Route (memory line only, prices Sep 14, 2026) | Memory cost | CPU-bound gaming against DDR5 on the same CPU | Capacity | What the memory does next |
|---|---|---|---|---|
| LGA1700 with DDR4-3200, 32GB | $190 | -9% to -14% against DDR5-7200 (Tom’s), about -17% against DDR5-6000 (TechSpot) | 32GB | Ends with the socket |
| LGA1700 with DDR5-6000, 32GB | $409 | baseline | 32GB | Carries to a later DDR5 platform |
| DDR5 platform with DDR5-6000, 16GB | $239 | one 16GB DDR5-7200 module beat 32GB of dual-channel DDR4 in Tom’s test; slower DDR5 narrows that | 16GB | One 16GB module: add a second later; 2x8GB: replace both to grow |
The budget route pays best when the graphics card is the limit, when capacity matters more than frame rate, or when the board is already owned, and least on a new high-end processor, where DDR5 lets it do more. Prebuilt machines, bundles and the timing of the next platform generation are the business of buy RAM now or wait.
If you already own the DDR4 machine, the drop-in beats the rebuild
This is the case the evidence favors most, and the reviewers most critical of DDR4 mostly agree with it. Tom’s Hardware’s June re-review of the $349 5800X3D concluded that it “really only makes sense if you already have an AM4 motherboard and DDR4 memory”, and TechSpot’s June buying guide said the same for both sockets.
Price an AM4 owner’s two options with September’s figures. B is your quote for an AM5 board and R is what the old processor, board and DDR4 would fetch.
drop-in Ryzen 7 5800X3D $349
switch, 16GB Ryzen 5 7600X3D 230 + DDR5-6000 16GB kit 239 + B $469 + B
extra cost 469 + B - R - 349 $120 + B - R
switch, 32GB 230 + DDR5-6000 32GB kit 409 + B $639 + B
extra cost 639 + B - R - 349 $290 + B - R
gain 7600X3D against 5800X3D in games about 20%
per point, 16GB (120 + B - R) / 20
The gain comes from Tom’s Hardware’s July article, which found the Ryzen 7 7700X3D “nearly 20% faster in average gaming performance” than the 5800X3D and said “The $230 Ryzen 5 7600X3D shows similar gains”.
| AM5 board quote B (illustrative, not a price) | Extra cost with R = 0 | Per percentage point |
|---|---|---|
| $100 | $220 | $11.0 |
| $150 | $270 | $13.5 |
| $200 | $320 | $16.0 |
One dated reference point for B exists: Tom’s Hardware’s best-motherboards guide, last updated on August 4, 2026, listed its B650 pick for AM5, the Gigabyte B650 Aorus Elite AX Ice, at $149.99, and warned that “B650 has been discontinued”; at B = $150 and R = 0 that is the middle row. Resale R reduces every row, and moving 32GB of DDR4 to 32GB of DDR5 instead of dropping to 16GB adds $170 to each.
Should you sell the DDR4 now, while it is expensive? Retail DDR4 is selling below module spot, so a used kit’s price is being set against cheap old stock, not replacement cost; eRacks’ tracking showed the median DDR4 desktop gigabyte falling 13% between March and September 2026 ($7.19 to $6.22) while DDR5 desktop rose 20%. At a fixed capacity the arithmetic is clear: even if a used 32GB DDR4 kit sold for the full $190 of the best new one, 32GB of DDR5 at $409 would still cost $219 more. Selling DDR4 to fund DDR5 is a bet on inventory, not a free trade. The live medians for the sum are on 32GB DDR4 and 32GB DDR5, and the part index shows what an exact kit is worth.
Adding DDR4 capacity: fill the slots once, and read what four modules cost
Three constraints specific to DDR4 decide an upgrade. The first is the module ceiling. DDR4 unbuffered modules stop at 32GB, so four slots stop at 128GB, and a two-slot board at 64GB. There is no 48GB DDR4 module to fill the gap the way there is on DDR5.
The second is what four modules do to speed, on AMD’s 5800X3D figures:
dual-channel peak = data rate x 8 bytes x 2 channels
DDR4-3200 3200 x 16 = 51.2 GB/s two modules, AMD's figure
DDR4-2933 2933 x 16 = 46.9 GB/s (-8.3%) four single-rank modules
DDR4-2667 2667 x 16 = 42.7 GB/s (-16.7%) four dual-rank modules
Those are official speeds before any XMP profile, and whether a 16GB DDR4 module is single- or dual-rank depends on its chips, as ranks and 3DS modules explains. AMD’s 5800X3D page and Intel’s ARK entry for the i5-14600K both stop at DDR4-3200, so anything faster is an XMP profile the board may or may not hold. The premium for 3600 is small, 307 / 281 = 9% more for a 2x16GB kit on the PCPartPicker averages, and a 3600 CL18 kit has the same first-word latency as 3200 CL16, 10.0 ns, with 12.5% more peak bandwidth. For a new DDR4 purchase the default is DDR4-3200, the official speed, with nothing left to prove; pay the 9% for 3600 only on a board whose memory support list includes it, and test it inside the return window. How timings and speed trade off is in the CAS latency guide.
The third is matching. Modules bought separately were never validated together, so Tom’s Hardware advised the same brand and specification, tested at JEDEC speed before the profile is enabled; with the high-bin DDR4 chips no longer made, the part number is the matching tool.
| Path from 2x8GB | Result | Modules and ranks | Official speed on a 5800X3D | What to compare on this site |
|---|---|---|---|---|
| Add a second 2x8GB kit | 32GB | four, single-rank if the 8GB modules use 8Gb chips | 2933, or 2667 if dual-rank | 8GB DDR4 modules, matched by part number |
| Replace with 2x16GB | 32GB | two modules | 3200 | 16GB DDR4 modules, against what the old pair sells for |
| Add 2x16GB to the old pair | 48GB | four, mixed sizes | 2933 or 2667, by rank | The least predictable; test before the return window closes |
What each path costs, before whatever the old pair fetches:
second 16GB kit Tom's index, DDR4-3200 16GB, Sep 14, 2026 $99
2x16GB, replace or add Tom's index, DDR4-3200 32GB $190
64GB as two 2x16GB kits PCPartPicker average, Aug 2026 2 x 281 = $562 $8.78/GB
64GB as one 2x32GB kit PCPartPicker average, Aug 2026 $614 $9.59/GB
premium for 32GB modules 614 - 562 = $52, or 9.59 / 8.78 = 9% more per GB
On the best-price index the gap is wider: its 64GB DDR4-3200 kit was $7.78 per gigabyte, 31% more than the $5.94 of its 32GB kit. Larger DDR4 modules cost more per gigabyte, so buying the largest modules is a premium paid for full speed and free slots, not a saving.
New DDR4 is a shrinking pool, so if you will want the capacity, buy it now and buy it once. On most boards the choice is then between two 32GB modules at full speed and that premium, or four 16GB ones at the four-module speed. Whether an old and a new kit will train together is in how to check what RAM fits.
Windows 10 ended in October 2025, and the oldest DDR4 machines are on borrowed time
An old DDR4 office PC or workstation bought to avoid DDR5 prices can run into Windows. Microsoft’s end-of-support page says Windows 10 reached end of support on October 14, 2025, and that consumer Extended Security Updates run until October 12, 2027. Enrollment is free if you sync your PC settings, costs 1,000 Microsoft Rewards points, or is a “One-time purchase of $30 USD or local currency equivalent plus applicable tax”.
Windows 11’s supported-processor lists start above the first DDR4 generations: Intel’s runs from 8th-generation Core, plus the X299 Core X-series, and AMD’s from the Zen+ 2000 series, such as the Ryzen 5 2600 and Ryzen 7 2700X. Both lists, re-read on September 26, 2026, against the DDR4 platforms:
| DDR4 platform | On Microsoft’s Windows 11 processor lists, Sep 26, 2026 |
|---|---|
| Intel 6th and 7th-generation Core (Skylake, Kaby Lake), mainstream desktop and laptop | No series entry found |
| Intel Core X-series 7000X (X299) | Yes |
| Intel Core i7 5800 to 6900 series (Haswell-E, Broadwell-E, X99) | No model found |
| Intel 8th to 14th-generation Core | Yes |
| Intel Xeon E5 v3 and v4 | No entry found |
| Intel Xeon Scalable, 1st to 3rd generation | Yes |
| Intel Xeon W-2100, W-2200, W-3100 | Yes |
| AMD Ryzen 1000 | No model found |
| AMD Ryzen 2000 to 5000 (not the 2200G and 2400G desktop APUs or 2000-series laptop chips) | Yes |
| AMD EPYC 7001 | No model found |
| AMD EPYC 7002 and 7003 (the 7003X Milan-X models are not listed) | Yes |
“Not listed” is a reading of Microsoft’s generated tables, not a statement that Windows 11 cannot be installed. A cheap Skylake or Kaby Lake desktop full of DDR4 is a supported Windows machine for about one more year, and only on Extended Security Updates; a used Xeon Scalable workstation is not bound by the same date. Check the processor on the CPU index and the machine on the server index before counting its DDR4 as a saving.
DDR4 laptops: about half the retail price per gigabyte and a 64GB ceiling
The laptop version of the budget route got a new product in August 2026: a variant of MSI’s Katana 15 HX C14, with up to a Core i9-14900HX and an RTX 5070, that takes DDR4-3200 SODIMMs rather than DDR5. Tom’s Hardware priced the difference:
32GB (2x16GB) DDR4-3200 SODIMM kit about $200 / 32 = $6.25/GB (Tom's Hardware, Aug 28, 2026)
32GB DDR5-6000 kit, type not stated over $400 / 32 = over $12.50/GB
largest DDR4 SODIMM 32GB, so two slots x 32GB = 64GB ceiling
The article did not say whether its DDR5 kit was a laptop kit, so the second line is a rough guide. It also doubted MSI’s “up to 96GB” figure, because no 48GB DDR4 SODIMM exists. A two-slot DDR4 laptop stops at 64GB. At the contract layer the gap disappears: TrendForce’s table for the second half of July 2026 priced a DDR4 16GB SODIMM at $265 and a DDR5 8GB SODIMM at $130, $16.56 against $16.25 per gigabyte. The retail laptop gap rests on stock, as the desktop one does.
On a used DDR4 laptop the upgrade worth most is often the one people skip: the empty second slot. A two-slot laptop with one module fitted runs its memory single-channel, and a laptop without a graphics card takes its graphics bandwidth from system memory; Intel’s specification footnotes say Iris Xe branding requires dual-channel memory. Adding a matching second module is a larger change than replacing the one module with a bigger one. Read the generation off the part number, because DDR4 and DDR5 SODIMMs look almost identical in a photograph.
Newer laptops are mostly not DDR4, and many cannot be upgraded at all: a SODIMM slot adds capacity, an LPCAMM2 module fills both channels on its own so an upgrade replaces it, and soldered memory is fixed at purchase. On a specification sheet, “SODIMM” or a slot count means socketed, “LPCAMM2” or “CAMM2” means one replaceable module, and “LPDDR5” or “LPDDR5X” with neither word means soldered. How to tell which one a machine has, and what fits it, is in how to check what RAM fits; the two socketed generations are listed here as DDR4 SODIMMs and DDR5 SODIMMs.
On the trackers, server DDR5 costs four to six times server DDR4 per gigabyte
For servers and homelabs the DDR4 case is strongest and the sources are weakest: there is no neutral price series for used server memory, only four sources that disagree on level and agree on direction.
| Source, date | Basis | DDR4 per GB | DDR5 per GB | DDR4 as share of DDR5 |
|---|---|---|---|---|
| eRacks, Sep 22, 2026 (a server builder’s own tracking) | Median of daily lowest street prices, ECC RDIMM | $5.56 | $31.05 | 0.18 |
| DatacenterDisk, Sep 26, 2026 | Median of Amazon US listings, 25 DDR4 and 16 DDR5 modules | $9.34 | $41.09 | 0.23 |
| DRAMeXchange, Sep 14, 2026 | Module spot, 32GB DDR5 RDIMM | not published | $64.06 | - |
| PCSP, page modified Sep 23, 2026 (a refurbished-server seller: its own refurbished DDR4 price against a new-market DDR5 spot quote) | 32GB RDIMM, refurbished DDR4 against new DDR5 | $5.00 to $7.34 | $12.50 to $15.63 | about 0.4 to 0.5 |
The two trackers put DDR5 at 31.05 / 5.56 = 5.6 times and 41.09 / 9.34 = 4.4 times DDR4; PCSP’s pairing puts the gap nearer two to two and a half times. Both eRacks and PCSP sell servers and neither method has been audited, so take the spread rather than one figure. A per-gigabyte ratio is not what a server costs, so take a worked host, a two-socket EPYC 7003 machine:
DDR4 host: two EPYC 7003 sockets, 8 channels each, one 32GB module per channel
16 modules x 32GB = 512GB
at $5.56/GB (eRacks) 512 x 5.56 = $2,847
at $9.34/GB (DatacenterDisk) 512 x 9.34 = $4,782
same 512GB in DDR5
at $31.05/GB 512 x 31.05 = $15,898
at $41.09/GB 512 x 41.09 = $21,038
at $64.06/GB (module spot) 512 x 64.06 = $32,799
The DDR5 lines flatter DDR5: the largest current DDR5 server processors, such as Intel’s Xeon 6980P and AMD’s EPYC 9965, have twelve channels per socket, so a balanced two-socket host at one 32GB module per channel holds 768GB, which at $31.05 is $23,846. Why filling only some channels gives up bandwidth is in server memory capacity limits.
What the cheap gigabyte buys less of is bandwidth. AMD lists the EPYC 7763 at DDR4-3200 on eight channels, 204.8 GB/s per socket (3200 x 8 bytes x 8), and the EPYC 9965 at DDR5-6400 on twelve channels, 614 GB/s (6400 x 8 x 12 = 614.4). The DDR4 socket has a third of the bandwidth. For a capacity-bound workload, a file server, a lab of idle virtual machines, a cache, that does not matter; for anything that streams memory it does. Running cost is the other term: every 100 W of continuous difference is 876 kWh a year. The live market is on registered DDR4, load-reduced DDR4, 32GB and 64GB modules, and which type a machine takes is on the server index.
A DDR4 RDIMM fits a desktop slot and will not start the machine
The homelab trap is specific to DDR4. DDR4 unbuffered and registered modules share the same key notch, so a server RDIMM seats in a desktop slot and the machine simply does not start; DDR5 moved the registered module’s notch, but DDR4 never did. Read the listing for the word “registered” rather than judging by the photograph. Workstations add a second check, the processor: Xeon W-2100 and W-2200 towers such as the Dell Precision 5820, HP Z4 G4 and Lenovo P520 take registered ECC only, but the same Dell and HP chassis sold with a Core X processor take non-ECC unbuffered modules (Dell’s owner’s manual; HP’s QuickSpecs, January 7, 2025; Lenovo’s PSREF, December 11, 2023). The full rules, including LRDIMM and 3DS modules, are in registered vs unbuffered memory and ECC vs non-ECC.
For cheap capacity at home there are three routes. A used server or workstation that takes RDIMMs, looked up on the Dell, HPE, Lenovo or Supermicro pages before buying the memory. An AM4 desktop with ECC unbuffered modules, where the board’s manual decides; the ECC DDR4 page includes registered modules, so confirm “unbuffered” on each listing. Or no ECC at all and the cheapest unbuffered gigabyte. Lots need reading module by module: Big Iron’s July 2026 homelab guide warned that “Broker lots pulled from decommissioned server fleets are frequently a mix of speed grades, ranks, and sometimes manufacturers even when sold as a matched set”.
Used memory comes from five places, and the server stream is starting to keep its own
DRAM does not wear out the way an SSD does: it has no write-endurance budget. What ages is the solder under each chip, which fatigues with thermal cycles, the spare rows consumed by post-package repair, and the edge contacts, as buying used RAM on eBay sets out. Which of those a module’s history makes likely depends on the route it took to market:
| Stream | Typical modules | What the history implies | Best evidence a seller can give |
|---|---|---|---|
| Decommissioned servers, through IT asset disposition and brokers | RDIMMs and LRDIMMs, bare, in lots | Constant load, watched by ECC, retired on a schedule | The source server model; ECC logs if kept; whether the set was pulled together |
| Employer or office disposals | Whole servers, workstations, desktops | As above, but handling unknown | A label photograph of each module |
| Desktop upgrade pulls | Enthusiast UDIMM kits with heat spreaders | Run on XMP by design, daily thermal cycles | The original kit part number; both modules together |
| Off-lease laptops | SODIMMs | Office duty, carried around | The machine model; a label photograph |
| Returns, open-box and “new (other)” | Anything | Includes the 2025-2026 return-fraud cases | A sealed box is not proof of contents, as shown below |
The first stream supplies most cheap registered memory, and it is the one that is changing. Meta described in June 2026 a CXL memory-expander chip that attaches DDR4 recovered from decommissioned servers to new DDR5-only EPYC servers, 256GB of DDR4-2400 beside 768GB of DDR5-6400 in each; The Register, reading the paper, reported that Meta’s servers serve three to five years, “but memory is useful for seven to ten years”.
Modules kept in service are modules that never reach a broker. That is an inference, not a measured flow, but the direction runs against the buyer, and demand points the same way: Apacer’s chief executive forecast in July 2026 that chip supply from the major makers to independent module makers could fall in 2027 to about 30% of its 2026 level. For DDR4 server memory the used stream is increasingly the market, and the server DDR4 page is where it shows.
Ex-server modules have the best history on the market, and one gap in it
Server modules run at constant temperature, which spares the solder, under ECC, and are retired on schedule rather than for a fault; the fleet studies are in buying used RAM on eBay. Age is not neutral, though: Google’s 2009 study found correctable-error rates starting to climb after 10 to 18 months in the field, and Facebook’s 2015 fleet study found that “older machines generally have higher failure rates than younger machines”. An ex-server module is a tested survivor, not a new part.
The gap is the error history. Google’s study found that 70-80% of uncorrectable errors were preceded by a correctable error in the same or the previous month, so a module’s correctable-error count is the best predictor there is of its next fault. That count lived in the host, not on the module, and it left with the host. Unless a seller kept the server’s logs, nobody can tell you what a module did in service; a long test after purchase is the substitute. What the label does tell you is roughly where a module came from. PCSP, a refurbished-server seller, ties DDR4 speed grades to server generations:
| Label speed | Platforms it came from (PCSP) | Channels per socket | Example chassis generation |
|---|---|---|---|
| 2133, 2400 | Xeon E5 v3 and v4 | 4 | Dell 13G, HPE Gen9 |
| 2666, 2933 | Xeon Scalable 1st and 2nd generation | 6 | Dell 14G, HPE Gen10 |
| 3200 | Xeon Scalable 3rd generation (Ice Lake); EPYC 7002 and 7003 | 8 | Dell 15G |
It is a seller’s summary: HPE runs EPYC 7002 at 2933 in a DL385 Gen10 and 3200 in a DL385 Gen10 Plus, same chip, different board. A label speed says which machines will take a module back at full speed. There is no DDR4-4800 server part, whatever a listing says: DDR4-4800 was sold only as overclocked desktop kits, and DDR4’s standard rates stop at 3200 MT/s. The live market for each grade is on DDR4-2400, DDR4-2666, DDR4-2933 and DDR4-3200, and the PC4 code on a server label is decoded in the speed guide.
“Ex-mining” means three workloads, and only one of them rewarded overclocking the memory
“Mining RAM” describes three histories, and what each asked of the memory is read from the workloads’ own documentation rather than from failure data. GPU mining was mostly Ethereum, and it ended at the Merge on September 15, 2022, when Ethereum moved to proof of stake; the graphics cards did the work and the system memory mostly sat there. Chia plotting is the opposite: the official BladeBit plotter says “416 GiB of RAM are required to run it” for in-memory CPU plotting, a sustained, write-heavy load. CPU mining with Monero’s RandomX needs about 2GB per NUMA node but is limited by memory bandwidth, “DDR4 memory is limited to about 4000-6000 H/s per channel (depending on frequency and timings)” per RandomX’s README, so it rewards overclocking the memory. Of the three, it is the one that plausibly ran the memory both flat out and overclocked.
| Workload | Memory it needed (primary source) | What it did to the memory | Likely module on resale (inference) | What to ask |
|---|---|---|---|---|
| GPU mining (Ethereum, to Sep 15, 2022) | Not documented; the GPUs did the work | Little memory load | Small desktop UDIMMs | Which coin, and when it stopped |
| Chia plotting (BladeBit) | 416 GiB (CPU mode), 256GB (GPU mode) | Sustained heavy writes | High-capacity RDIMMs | Plotting or only farming, and for how long |
| CPU mining (Monero RandomX) | 2080 MiB per NUMA node | Channels saturated day and night; often overclocked | Desktop kits, sometimes high-bin | Whether the memory was overclocked |
Temperature is the usual worry, and the evidence is narrow: Google’s fleet study found only a marginal effect on error rates within the roughly 20 °C range of normal data-center operation, which says nothing about an unventilated mining frame. Ask which of the three the seller means; Chia plotting memory is likely to be registered, which changes what to test for.
Every enthusiast kit is rated as an overclock, and no module remembers how it was run
A DDR4-3600 kit reaches 3600 only through its XMP profile. Without it the kit boots at the JEDEC grade stored in its SPD, commonly DDR4-2133 or DDR4-2400. G.Skill began paying a $2.4 million US class-action settlement in August 2026 over exactly this, agreeing to state on packaging and listings that XMP or EXPO overclocking is needed to reach rated speeds; it denied the allegations, as Team Group did in its $1.1 million settlement over the same claim in May. So a used enthusiast kit ran overclocked by design if its owner enabled the profile, and that is normal use, not abuse.
The question that matters is whether it was pushed beyond its profile, or run at raised voltage, and no DDR4 module can answer it: DDR4 takes its supply from the motherboard’s regulator, so the voltage history lived in the board. What a buyer can check:
- the part number on the label against the maker’s rated profile for that kit;
- the SPD contents against the label, since G.Skill’s warranty excludes modules with modified SPD firmware;
- stability at the XMP profile, then at JEDEC speed, inside the return window.
All three are explained in buying used RAM on eBay, XMP as a memory-controller overclock in the speed guide, and the most-tuned grade trades on DDR4-3600.
“New” is not a provenance either
In 2023, during a DDR4 glut, TrendForce reported that DDR4-3200 chips were being desoldered from old server modules, re-flashed and built into new consumer kits; in 2026 recovered silicon is traded again, with TrendForce’s September 16 spot report noting that “reballed chips sustained the upward price momentum seen last week”. The reported cases of such chips in consumer kits date from 2023, and the incentive is larger now that DDR4 silicon costs more than DDR5’s. Lower grades are the other risk: DRAMeXchange quotes a separate DDR4 line labeled eTT at $13.65 for a 16Gb chip on September 24 against $83.78 for the standard 3200 grade. A chip at a sixth of the standard price is not the same part, whatever the module’s label says. A cheap new no-name kit is not evidence of first-grade silicon. The rest are 2025-2026 cases, most of them DDR5, whose lessons carry to DDR4:
| What the listing says | What can be inside (source, date) | Where it gets caught |
|---|---|---|
| New, no-name DDR4 kit | Recovered server chips (TrendForce, 2023); lower-grade eTT chips (DRAMeXchange lists the grade, 2026) | Package markings and date codes against the kit’s own date; SPD maker ID |
| New, sealed retail box | Two DDR or DDR2-era modules and a ballast plate in a DDR5 box, sold through Amazon in Spain (Tom’s Hardware, Dec 17, 2025) | Opening on camera; the capacity the machine reports |
| Bare SODIMM, sold as-is or untested | Plastic or fiberglass dummy chips, SK hynix markings under a Samsung label, on Japanese second-hand marketplaces (Tom’s Hardware, May 11, 2026) | Label maker against package maker; testing |
| RGB module | A legitimate 0GB dummy, about $44 to $57 in the US, or a “1+1” pack with one dummy (Tom’s Hardware, Sep 4, 2026) | Capacity reported by the firmware, not appearance |
A label maker that differs from the chip maker is a flag to test, not proof, because some genuine modules carry another maker’s chips. And a bare module is normal for SODIMMs and server RDIMMs; if anything, Tom’s Hardware noted, fakes are easier to spot on bare modules.
Between a cheap new no-name DDR4 kit and a used branded one at a similar price, the used kit is usually the more checkable purchase. Its part number names a rated profile and a chip maker, and the label, the packages and the SPD can each be tested against them; the no-name kit’s label promises nothing a test can contradict. The part index shows what a part number should be.
The warranty stays with the first owner, and G.Skill’s “lifetime” ends with the model
In the US, G.Skill’s and Corsair’s terms do not pass the warranty to a second owner.
| Maker | Transferable to a second owner | What “lifetime” means | Notable exclusions |
|---|---|---|---|
| G.Skill (read Sep 26, 2026) | No; receipts from online auction sites, liquidators and clearance houses are not valid proof | “the period during which the specific model number is actively manufactured and supported”; ends at end-of-life | Modified SPD firmware; removed or altered labels |
| Corsair (edited Aug 11, 2026) | No; in the US, DRAM is covered only if bought from Corsair or an authorized reseller | Limited lifetime; a discontinued product is replaced with one of similar function and capacity | Crypto-mining damage; improper voltage |
| Crucial (Micron, winding down) | Not stated | - | Where no replacement exists, the maximum is a refund of the original price, paid in gift or prepaid cards |
G.Skill’s “lifetime” warranty ends when G.Skill discontinues the model, so on an older DDR4 kit it may already be over. And Crucial is gone from retail: Micron announced on December 3, 2025 that it would ship Crucial consumer products only until the end of February 2026, and Crucial’s warranty page now says that a refund of the original price, taxes and shipping is “the maximum resolution available during the closure”. Tom’s Hardware reported in August 2026 that Micron first offered one claimant $241.86 for a 48GB DDR5 kit when the cheapest 48GB kit on sale, a faster one, cost $649.99, 37.2% of that price, before sending three 16GB modules instead.
So a used kit bought in the US normally carries no maker’s warranty. What remains is eBay’s Money Back Guarantee, whatever the seller’s return policy; the steps and deadlines are in buying used RAM on eBay.
Ex-server hardware from a refurbisher is the partial exception, because the seller’s terms replace the maker’s: PCSP’s June 22, 2026 buying guide offers returns within 90 days and a server warranty of “1 year, extendable up to five”, with a 20% restocking fee on items “returned as defective but found fully working”. Ask any seller for the same two things, a window long enough for a full test and warranty terms in writing.
One US price site’s rule of thumb is to buy used only at about 60-70% of a comparable new kit. A sum fits it to your own case:
break-even chance of failure after the return window = discount / replacement cost
new 32GB DDR4-3200 kit, best price (Tom's index, Sep 14, 2026) $190
used kit at, for example, $150 discount $40
break-even 40 / 190 = 21%
Unless you believe a used kit that passes its tests has more than a one-in-five chance of failing after the window closes, the used kit and its $40 saving are the better bet. The probability is your own input.
Filter by the machine first: which page on this site answers which question
Every listing page here sorts by price per gigabyte, prints the cheapest and the
median at the top, and compares Buy It Now prices by default, because a bid is not a
price (?bids=show adds bid-only auctions, ?auctions=no removes auctions entirely).
Every module listed is sold new, open-box, refurbished or used, with its condition
shown on each row, and the part index groups every listing of one part
number.
| Question in this guide | Page | What to read |
|---|---|---|
| Is DDR4 still about half DDR5 per gigabyte? | 32GB DDR4 against 32GB DDR5 | The two medians |
| An AM4 or LGA1700 kit | DDR4 desktop, DDR4-3200, DDR4-3600 | The median, then the part number |
| Filling four slots | 64GB DDR4 in total | Kits against two separate pairs |
| Matching a second kit | The part index | Same part number, both modules |
| A laptop module | DDR4 SODIMMs, DDR5 SODIMMs | The generation, from the part number |
| Server capacity | Registered, load-reduced, by machine | The machine’s supported type first |
| ECC for AM4 | ECC DDR4 | Confirm “unbuffered” on each listing |
| Every generation at once | All generations | Cheapest and median price per gigabyte |
| How prices have moved | Trends | The like-for-like line for each generation |
A DDR4 desktop row far below its page’s median is worth a look at the label first: it may be a server module, one module from a kit-priced title, or a typo.