A 48GB kit from KingBank using 24Gb DDR5 dies manufactured by China's ChangXin Memory Technologies (CXMT) achieved 8600 MT/s operation on an ASUS ROG Crosshair X870E Apex. Hardware outlet Uniko's Hardware published the verification results on July 15, provided by ASUS overclocker SafeDisk. This is a case that far exceeds the DDR5-5600 spec AMD sets for the Ryzen 7 9800X3D, and it marks a step forward in platform verification of CXMT DDR5 at the 8000 MT/s class. However, the report also notes that gains become harder to achieve even with increased voltage, and tightening timings is difficult. Reaching 8600 MT/s and being able to reproduce that same performance in mass-produced kits are two separate matters that need to be considered independently.
A 6000 CL36 48GB Kit Pushed to 8600 CL44
The kit used in testing was a two-module KingBank set using CXMT's 24Gb dies—that is, 3GB per die. The rated profile is DDR5-6000, 36-38-38-88, at 1.25V, with each module at 24GB for a total of 48GB. The CPU was a Ryzen 7 9800X3D, and the motherboard was the two-slot ROG Crosshair X870E Apex.
In the published images, the actual clock reached 4309.4MHz, translating to roughly 8619 MT/s in DDR terms. The manual settings were 44-56-56-126 with a Command Rate of 1T. RunMemtestPro ran for 58 minutes and 32 seconds, recording an average coverage of 110.19%. At minimum, this particular unit under this test environment was able to complete a full memory test pass at the 8600 MT/s class.
On the other hand, this is not a verification of a warrantied DDR5-8600 product. The BIOS shown in the image is version "0802" dated July 13, which differs in number from the beta BIOS 2402 that ASUS made publicly available on July 15. The roughly 59-minute test also does not guarantee long-term operation or stability across other units. What SafeDisk disclosed were the results achieved and observations, but the voltage range tested, temperatures, and the settings used for the SK hynix kit used as comparison were not disclosed.
The Number 8600 Alone Doesn't Measure Effective Latency
MT/s indicates the number of data transfers per second. While higher frequencies increase theoretical bandwidth, the time it takes to begin reading data requested by the CPU also depends on timings such as CL, tRCD, and tRP. Converting clock cycles into real time reveals what this particular setting improved and what it left unchanged.
| Setting | CL (real time) | tRCD / tRP (real time) | tRAS (real time) |
|---|---|---|---|
| DDR5-6000 36-38-38-88 | 12.00ns | 12.67ns | 29.33ns |
| DDR5-8600 44-56-56-126 | 10.23ns | 13.02ns | 29.30ns |
The conversion formula is "timing cycle count × 2000 ÷ data rate." CL shortened from 12.00ns to 10.23ns. However, tRCD and tRP slightly lengthened from 12.67ns to 13.02ns, and tRAS remained essentially unchanged. In other words, in order to achieve a higher data rate, the cycle counts were relaxed—so while bandwidth and CAS latency improved, the time required for row selection and precharge did not shrink.
As part of SafeDisk's observations, an assessment was also published stating that at the same clock speed, this memory is slower than SK hynix DRAM. However, bandwidth, latency, and real-world application figures measured on the same CPU, same motherboard, and same clock speed have not been disclosed. Secondary and tertiary timings, as well as the CPU's memory controller ratio, also affect results, so it's not possible at this stage to generalize the difference as a specific percentage. What can be confirmed from the table above is that when this particular CXMT kit was pushed from 6000 MT/s to 8600 MT/s, the major latencies other than CAS barely improved.
The 8000 MT/s Wall Pushed Higher by BIOS Updates
CXMT has stated regarding its own DDR5 that it offers both 16Gb and 24Gb variants, and that the maximum data rate of its chips has exceeded 8000Mbps. On top of this, adjustments from various motherboard manufacturers have been layered in. On July 15, ASUS released beta BIOS 2402 for the ROG Crosshair X870E Apex, noting in the update details that it improves compatibility with CXMT chips, memory performance, and system stability.
MSI also published verification results for the Intel 800 series on July 14. On the two-slot MEG Z890 UNIFY-X, a 48GB configuration using CXMT's 24Gb chips exceeded 100% coverage in MemTest at DDR5-8600, 46-56-56-134. On the four-slot PRO Z890-S WiFi as well, a 32GB configuration passed the same benchmark at DDR5-8200, 44-56-56-132.
According to MSI, for the Intel 800 series, they implemented BIOS memory training and timing adjustments specifically for CXMT. Meanwhile, the company also explained that on AMD platforms, dedicated tuning allowed them to surpass the previous 6800 MT/s ceiling that had existed for CXMT DDR5. This reflects progress in the process of aligning signal conditions with DRAM characteristics at boot time. The frequency ceiling is not determined by the DRAM die alone—it is the result of the board's trace routing, the number of DIMM slots, the CPU's integrated memory controller, and the BIOS all working together.
Even the manufacturers themselves attach conditions to reproducibility. MSI explicitly states that overclocking results vary depending on the quality of the CPU's memory controller, the module and IC batch, and the system configuration. The Ryzen 7 9800X3D's official maximum memory speed is DDR5-5600, meaning 8600 MT/s falls under manual tuning outside the warrantied range. ASUS's published QVL similarly cautions that platform constraints may prevent operation at maximum speed.
Can It Be Turned Into a Retail Product Despite Batch Variance?
What's tricky about this report is that SafeDisk assessed the silicon variance between batches as significant. If gains in frequency and timing become difficult to achieve even with increased voltage, and if the appropriate settings vary widely from lot to lot, module manufacturers will need to either increase binning or set more conservative published profiles. Even if a single kit reaches 8600 MT/s for record purposes, that doesn't mean a user who buys the same model number will be able to use the same settings.
As of the confirmation date of July 16, 2026, the names KingBank and CXMT do not appear in the publicly available QVL for the ROG Crosshair X870E Apex. While beta BIOS 2402 touts improved CXMT compatibility, a list of supported kits and guaranteed speeds has not yet been provided. What users can currently rely on are the EXPO or XMP profiles listed on the product itself and the QVL published by the motherboard manufacturer. This is not yet the stage to purchase a DDR5-6000 CL36 kit in the expectation that it can be pushed to 8600 MT/s.
Still, the fact that more platforms are becoming capable of handling CXMT DDR5 at the 8000 MT/s class carries significant weight. In its June earnings materials, Micron projected that demand for DRAM and NAND would significantly outpace supply, with the shortage continuing beyond 2027. If the number of suppliers for PC-oriented DRAM increases, dependence on the big three manufacturers can be reduced. However, expanding supply volume and controlling variance in high-speed products are separate challenges.
The next benchmark for judgment isn't going to be an even higher single record. It will be whether the same profile can be reproduced across public BIOS releases and multiple production batches, and whether bandwidth, effective latency, and application performance can be compared against SK hynix kits under identical conditions. Once all of that is in place, CXMT's achievement of exceeding 8000 MT/s will transform from a laboratory milestone into genuine retail-product performance.