XMP, EXPO and why your memory runs slower than the box says

There is a further distinction that trips up even careful readers, and it concerns what JEDEC has actually standardised at the top end.

In October 2025 JEDEC updated the SPD contents standard so that it "adds support for memory modules executing up to DDR5-9200 speeds" 5. Separately, the DRAM standard itself, JESD79-5C, brought an "Expansion of timing parameters definition from 6800 Mbps to 8800 Mbps" 6.

Those are two different things and the difference matters. SPD gaining the vocabulary to record a DDR5-9200 rating is not the same as JEDEC defining a DDR5-9200 DRAM speed grade with timings. A 9200 MT/s kit is an overclocking product whose rating the standard can now describe. It is not a standard-speed part.

For reference, JEDEC's own account of where DDR5 started: it "is expected to be launched at 4.8 Gbps (50% higher than DDR4's end of life speed of 3.2 Gbps)" 7. The current DRAM standard is JESD79-5D, version 1.41, published November 2025 8.

One limit of this page, stated rather than glossed: JEDEC’s speed-bin tables are paywalled — JESD79-5D is free to paying JEDEC members on login and costs US$423 to everyone else 8 — so the timings quoted further down come from manufacturers’ datasheets rather than from the standard document directly 1.

The clearest historical illustration of the two-regions idea comes from Intel's own DDR3-era white paper, which simply names the byte range its profiles occupied: "XMP DDR3 DIMMs use bytes 176–255 to program specific profiles." 4

Do not carry those byte numbers into DDR5. DDR5's SPD is the 1024-byte device described above with a 640-byte JEDEC region 11, and no official statement of the byte offset at which a DDR5 XMP block begins was found in this check — not in Intel's published XMP material, not in JEDEC's free documents. The structural point stands on the Micron datasheet; the specific DDR5 offset is unverified here and this page does not assert one.

A related gap in the same direction. Intel publishes a document numbered 852792, which its own content-details page titles Intel Extreme Memory Profile (Intel XMP) 3.0 Memory for Intel Core Processors Datasheet and describes as providing "Intel® Extreme Memory Profile 3.0 (Intel® XMP) DDR5 unbuffered DIMM memory supplier self-certifications" 13. The XMP landing page links the same file under the shorter label "Intel® XMP 3.0 for DDR5 Memory Datasheet" 2. It is a freely downloadable PDF, not a gated one; it was not opened for this check, and by Intel’s own description it is a set of supplier self-certifications rather than a structure specification. No Intel document defining the XMP 3.0 data structure was found in the material read here, which is why so much of what circulates about it comes from third parties.

One consequence for anyone diagnosing a machine: because the JEDEC region is what a system reads first, a module whose profile fails to apply will still boot, silently, at the standard speed. A build that feels slower than it should is far more often an unapplied profile than a defective part.

A contradiction on Intel's own pages, recorded because a careful reader will hit it. Intel's configuration article maps the versions to processor generations like this: "Intel® XMP 2.0: For DDR4 and Intel® Core™ 12th and 13th Generations Processors" and "Intel® XMP 3.0: For DDR5 and Intel® Core™ Generations Processors older than the 13th and 12th Generations" 15. The second statement cannot be right: DDR5 and XMP 3.0 arrived with 12th Gen, so processors older than that support no DDR5 at all. Intel’s dedicated article on the DDR5 question, meanwhile, says only that "Intel XMP 3.0 technology is exclusive to DDR5 memory modules" and that "For DDR4 memory modules, use Intel XMP 2.0 technology" 14 — it gives no generation mapping at all. Ignore the configuration article’s generation mapping: Intel publishes no reliable one.

No voltage range for XMP profiles appeared in any Intel material read for this check. It documents that XMP 3.0 controls module voltage and names the three rails 2 16, but no numeric range appeared in anything read for this check. The only figure available is from technical press, and is labelled as such: one analysis reports that the module's power-management chip on XMP parts "supports additional capabilities of Intel XMP requirements and overclocking capabilities which are out of JEDEC spec", including voltages "higher than 1.435V with 5/10mV switches" 17. Standard DDR5 runs at 1.1 V 18.

Two practical notes from Intel's own checklists. Before expecting a profile to work: "Check if the motherboard is listed as compatible with Intel® XMP by confirming it with the manufacturer" and "Check if the memory module is Intel® XMP Certified" 19. And Intel does not document the BIOS steps itself — it says "To find the right BIOS menu or tuning utility for your system, contact the motherboard vendor or your OEM for the related support." 15

Finally, Intel's own acknowledgement that a profile is not always the right setting to be in. Its Dynamic Memory Boost feature exists because "Rather than running aggressive settings all the time, its new Dynamic Memory Boost Technology intelligently switches between the baseline (default) speed and XMP profile of your choice whenever it detects a performance-bound workload." 16

A limit of this page that is a limit of AMD's site. AMD's memory compatibility list did not render its contents in this check. Two retrievals of it returned only navigation, a banner, the page title and the one descriptive sentence quoted above — no kit rows, no column headings, no filters, no pagination and no download link 24. The data is evidently injected by script. So this page cannot say what is on the list, how many kits it holds, or whether it carries a disclaimer. Nothing here should be read as a claim that any of those is absent.

On the cross-platform question — an EXPO kit in an Intel board, or an XMP kit in an AMD one — neither AMD nor Intel was found to publish an answer. The most direct guidance comes from a module maker, and it is worth quoting in the order it appears, because the first sentence alone would mislead:

Corsair writes that "If you buy DDR5 RAM that supports Intel XMP 3.0, that memory won't work with an AMD platform. You're locked into Intel's chips unless you're willing to buy more memory specifically for a platform change." A little further down the same article it says: "In practice motherboard manufacturers have done a good job supporting both standards—XMP on AMD platforms and EXPO on some Intel ones." 26 Its separate article on the question is blunter still: an XMP kit in an AMD board means "In practice there's a good chance that the memory will work, albeit with a few caveats", and "You should be able to get it working, just not at a fast speed setting" 27.

Some brands now sidestep the problem. Corsair states that "Since this page was originally published, all CORSAIR AMD Memory Kits are now dual profile, meaning that they are compatible with EXPO and XMP. However, Intel Memory kits are not dual-profile." 26 And two brands publish a way to tell which standard a kit carries before buying: Corsair notes that "Intel XMP kits will be denoted with C__, AMD EXPO will be Z__" in its DDR5 part numbers, giving "CMP32GX5M2X7200C34 = Intel XMP" and "CMP32GX5M2B6000Z30 = AMD EXPO" as its own examples 28; Kingston encodes it as a "Profile Type" field where "E - AMD EXPO & Intel XMP" 29. Kingston’s blank value covers XMP-only, Plug N Play and dual kits alike, so an explicit E tells you something but a blank does not 29.

AMD’s fullest statement is in its Threadripper and Ryzen Master overclocking guide — note the subject of that document — and it is unambiguous enough to quote at length. The guide states that "AMD does not provide support or service for issues or damages related to use of an AMD product outside of the Specifications or outside of factory settings", that "Recipient understands that operation of the product outside of AMD's published specifications will void any AMD warranty", and — in capitals in the original — "DAMAGES CAUSED BY USE OF YOUR AMD PROCESSOR OUTSIDE OF OFFICIAL AMD SPECIFICATIONS OR OUTSIDE OF FACTORY SETTINGS ARE NOT COVERED UNDER ANY AMD PRODUCT WARRANTY." 31

Intel’s equivalent long-form warning is older, and is still published in Intel’s own document library. Its last sentence is the bluntest of the statements read here: "Warning: Altering clock frequency and/or voltage may (i) reduce system stability and useful life of the system and processor; (ii) cause the processor and other system components to fail; (iii) cause reductions in system performance; (iv) cause additional heat or other damage; and (v) affect system data integrity. Intel has not tested, and does not warranty, the operation of the processor beyond its specifications." 4

Two scoping points, because this is the section most likely to be quoted out of context. First, watch the scope, because the two companies do not draw it in the same place. Intel’s XMP-page sentence names the processor warranty specifically 2; its general boilerplate is broader, covering "any product warranties" and "the processor and other components" 15. AMD’s GD-106 is broader still: it names "AMD processors and memory", voids "any applicable AMD product warranty", and adds that overclocking "may also void warranties offered by the system manufacturer or retailer" 20. What neither clause governs is your memory kit’s own warranty, which comes from the module brand — those are covered on the individual brand guides in this section, where "lifetime" turns out to mean different things at different brands. Second, the Intel white-paper warning dates from 2009 4, so present it as Intel's long-standing position rather than as current DDR5-era text.

What neither company publishes, and this page therefore does not claim: any statement that running a profile within its rated settings is covered. Intel's conditional "beyond the given specifications" implies a line exists; nothing read here says where it is.

There is a further mechanism by which a higher data rate can cost you performance, and it lives in the memory controller's own clock ratio. Intel defines its version: "Gear 1 means processor memory controller and memory speed are equal. Gear 2 means processor memory controller operates at half the memory speed (such as CPU memory controller is at 1600MHz while memory speed is at 3200MHz when operating as Gear 2)." 36

Note carefully what Intel does not say there. That article defines the ratio and says nothing about which gear gives lower latency or better performance, and it is written about DDR4 and 11th Gen Core processors 36. The widely-repeated claim that Gear 1 is better for latency is not Intel's, and this page does not attribute it to Intel.

AMD exposes analogous clocks under different names, though no AMD statement equating them to Intel’s gears was found. Its Threadripper and Ryzen Master overclocking guide glosses them as MEMCLK ("Internal and external memory clock"), FCLK ("Data Fabric clock – equal to Memory clock") and UCLK ("Internal Memory Controller clock – equal to Memory clock") 31. AMD does not use the word "Gear" in anything read for this check, and no AMD statement on 1:1 versus 1:2 UCLK-to-MEMCLK ratios was found.

The ratio framing applied to AMD comes from a module brand, borrowing Intel's vocabulary, and should be attributed there rather than to AMD: Kingston describes Gear 1 as the memory controller and memory running at the same frequency, "Best for gaming or latency-sensitive workloads", with DDR5-6000 "generally the highest speed that can achieve Gear 1 out of the box", and Gear 2 as the controller at half the memory frequency, "Best if you need memory bandwidth for AI, video editing, 3D rendering or any applications where throughput is more important than responsiveness" 37. The much-quoted "DDR5-6000 sweet spot" for AMD builds is Kingston’s figure, not AMD’s. AMD does publish DDR5-6000 test configurations of its own on the EXPO page 20, but it never calls the figure a sweet spot, and no AMD statement on memory-controller ratios was found here.

Sourcing note. JEDEC’s speed-bin tables are paywalled. JESD79-5D (version 1.41, November 2025) is free to paying JEDEC members on login and costs US$423 to everyone else 8, so the figures above come from manufacturers’ datasheets rather than from the standard document 1 18. No complete A/B/C bin list was verified against a single authoritative table in this check, and an early retrieval returned per-grade CL figures that conflicted with the module datasheet on a second reading; they were discarded as extraction error rather than published. That is why this page gives one fully worked grade and no table of them.

There is one more reason an unstable fast profile can be hard to spot, and it comes from DDR5 itself. Every DDR5 chip corrects some of its own errors. JEDEC frames this as a manufacturing enabler rather than a user feature: "On-die ECC and other scaling features enable manufacturing on advanced process nodes." 7

What on-die ECC does not do is the part that matters here, and the clearest statements come from a module manufacturer and from technical press rather than from JEDEC’s free material: on-die ECC "corrects single-bit errors inside the DRAM chip before data is sent to the CPU", but "does not protect data moving between the memory module and the CPU or GPU, and cannot correct errors that occur outside the chip, such as on the module traces or at the memory controller", and "does not turn a standard DDR5 module into a server-grade ECC module" 38. The mechanism, per EE Times: on-die single error correction ECC "pairs 128 data bits with 8 parity bits to form a 136-bit codeword stored in the DRAM during a write command", and "On-die single-bit ECC does not eliminate the need for error correction across the whole memory subsystem." 39

Put those together and the practical warning writes itself. On-die ECC does not cover the module traces or the memory controller 38 — which, on our reading rather than ATP’s, is exactly where an aggressive data rate bites hardest. An over-ambitious profile need not crash; it can corrupt quietly. No source read for this check frames on-die ECC as protection against overclocking instability, and this page does not either.

Now the fact most likely to change a purchase: a CUDIMM does not behave the same way on Intel and AMD.

Corsair states it plainly: "CUDIMM modules can run in bypass mode or in single/dual PLL mode. Currently, on AMD platforms only the bypass mode is supported." 46 And on its technical page: "AMD doesn't support the Single or Dual PLL modes yet. Although if you plug a CUDIMM into an AMD platform, they tolerate the use of it by operating in Bypass mode only." Alongside: "As of now, CUDIMMs are fully compatible with Intel platforms." 45

In bypass mode the clock chip is effectively switched out of circuit — "the CKD doesn't regenerate and amplify the signal; basically, it works as a traditional UDIMM" 45. Which means the feature you paid for is doing nothing. On Intel's side the same page reports that "on Intel® Z890 motherboards, the memory speed can go over 9,000MT/s" 45.

Two honest caveats on that finding. First, Corsair’s own bypass-mode ceiling is inconsistent — and note that both figures are stated about older DDR5 platforms, not about AMD. Its technical explainer says bypass mode "is used for compatibility on older DDR5 platforms. However, speeds may be limited to 6,000MT/s" — note the hedge 45. Its support article for chipsets earlier than Z890 says instead that "the base speeds are lower than 6400MHz depending on the platform" 47 — Corsair’s "MHz" there, reproduced as written, for what is a MT/s figure. Corsair publishes no single consistent number, this page reports both rather than picking one, and neither of them is an AMD data-rate ceiling. Second, and more importantly: no AMD or JEDEC source was found that mentions bypass mode, PLL modes, or an AMD data-rate limit at all. Of the AMD pages readable for this check, none mentions CUDIMM or the CKD 20 23; AMD’s memory compatibility list is script-rendered and its contents could not be read either way 24. And JEDEC's free material describes the clock driver without addressing vendor support 41 42. The asymmetry rests entirely on one module maker's technical documentation.

The practical reading for a builder is therefore: on a current Intel platform a CUDIMM is what unlocks the speeds above roughly 6400; on AM5 today it is, at best, an ordinary UDIMM that cost more. Verify against your own board's documentation before buying, because this is exactly the kind of fact that a firmware update or a new processor generation changes — and because, as the caveat above says, the companies whose silicon decides it have published nothing on the question.