AMD has detailed AMD MEXT, a predictive memory software that uses artificial intelligence to turn NVMe flash storage into an extension of DRAM — and, according to the company's own tests, keep up to 95% of performance with half the DRAM in EPYC servers.
The technology arrives as DRAM grows expensive and scarce: AMD cites a TrendForce forecast of a 10% to 15% rise in conventional DRAM contract prices in the fourth quarter of 2026. AMD MEXT's pitch is to make better use of the memory that already exists instead of buying more.
How AMD MEXT works
AMD MEXT moves rarely accessed memory pages — so-called "cold pages" — to flash, and uses an AI engine built on transformer-derived models to predict which pages the application will request next, moving them back to DRAM before they are requested. From the application and operating system point of view, the data is always in DRAM.
The solution is software-only: it needs no GPU and uses two CPU cores, one for the AI engine and one for the daemon. It runs on Linux distributions including RHEL, Rocky, AlmaLinux, Oracle Linux, SUSE, Debian and Ubuntu, and requires a dedicated NVMe drive with swap disabled. AMD acquired the MEXT startup in June 2026 and folded the technology into its data center portfolio.
The numbers AMD presented
In a presentation released on November 11 at the AMD In Depth event, company executives detailed results from AMD's own IT infrastructure:
- 95% of performance retained in a test with Synopsys PrimeTime, a chip timing analysis tool, with half the DRAM replaced by flash;
- 99% of performance retained in a test with Apache Spark, also with half the DRAM;
- 80% of performance when DRAM falls to 25% of memory and flash takes 75%;
- 145 servers already running MEXT inside AMD IT, with an estimated $4.9 million saved in avoided DRAM purchases;
- 2,000 servers is the rollout target for 2027.
An official AMD case study from October 30 gives the first phase's figures: 86 hosts, 58 TB of additional usable capacity and savings of about $2.78 million, or 82% of the cost of an equivalent all-DRAM expansion.
Why it matters now
DRAM scarcity and prices are squeezing data centers, cloud providers and AI companies, which need ever more memory for agents and intensive workloads. According to AMD, AMD MEXT can cut DRAM needs in EPYC servers by 20% to 75%, depending on the workload, and expanding memory with flash costs 83% less than an equivalent all-DRAM expansion.
The software will be licensed through OEM partners and works exclusively with AMD hardware. According to trade press reports, the company is also working with Hewlett Packard Enterprise (HPE) to bring the solution to enterprise customers, and offers the RAM View observability tool for free.
Limits and caveats
The performance figures come from AMD's own tests and have not been independently verified — results vary by workload, and AMD itself acknowledges that some workloads have not yet met performance expectations. AMD MEXT is a paid license, which factors into the total cost, and the technology only works on servers with AMD processors. The story connects to Geekiki's coverage of the 6th-generation EPYC "Venice", launched in July with up to 256 cores and a focus on AI agents.
Image credit: official AMD MEXT promotional image, via AMD (amd.com).
Sources
- AMD — AMD MEXT Predictive Memory Software
- AMD — AMD IT Avoids 82% of DRAM Expansion Costs with AMD MEXT Software
- The Register — AMD: Where we're going we don't need (as much) DRAM
- Notebookcheck — AMD's new trick lets servers use half as much DRAM with minimal performance loss
Transparency: this article was produced with the assistance of artificial intelligence, cross-checked against public posts and internet sources, and reviewed before publication. Please consult the original sources listed above to confirm the data.
By Muse
