AMD announced plans to replace a portion of server random-access memory with cheaper flash memory during a presentation in Las Vegas, Tweakers reported on November 10, 2026.
Rising memory prices have driven RAM to account for three-quarters of a server’s total hardware cost. The new hardware approach aims to lower these production and operational expenses for enterprise customers.
The software enabling this hardware shift is called MEXT, named after the firm AMD acquired in June 2026. MEXT trains an artificial intelligence model live to determine which memory pages require actual DRAM and which can reside in flash storage.
The artificial intelligence model learns a workload’s memory behavior within one second using a single CPU core. AMD avoided pretrained models because usage patterns vary significantly across different tasks, and enterprise clients demand that workload training stays private.
Flash memory runs five hundred times slower than DRAM, but it costs half as much and consumes thirty percent less power. Research into memory utilization shows that much of loaded data is rarely modified, allowing slower storage tiers to handle it without severe performance penalties.
System administrators can inspect how cold or hot specific memory regions are by using the RAM View tool included with the software package.
MEXT adds up to three times the original memory capacity using NVMe SSDs. Systems running flash storage achieve eighty to ninety percent of the performance seen in servers equipped entirely with standard working memory, according to AMD benchmarks.
AMD distributes MEXT as a Linux driver with explicit support for Red Hat Enterprise Linux, Oracle Linux, SUSE Linux Enterprise Server, and Ubuntu. The software operates as a paid feature requiring a license purchased exclusively through AMD’s enterprise sales division.
