Overview
- AMD publicly unveiled MEXT in November 2026 and says Phase 1 is ready for deployment on EPYC servers with NVMe drives.
- MEXT uses lightweight Linux drivers and an on-CPU AI engine that predicts which memory pages will be needed next and moves cold pages to local NVMe flash before the OS requests them.
- AMD and former Mext CEO Gary Smerdon report internal pilots across more than 100 hosts added about 58 TB of usable capacity at an estimated rollout cost of roughly $601,000 versus $3.3 million for equivalent DRAM, a vendor-claimed savings of about 82–83%.
- Vendor benchmarks claim MEXT can cut physical DRAM by roughly 20–75%—commonly around 50%—while retaining roughly 95–99% of baseline performance on tested workloads, but those results are from AMD/Mext and await broader third-party validation.
- If adopted at scale, MEXT could ease DRAM supply pressure and lower data center memory bills by using far cheaper, lower-power NVMe flash for cold pages, with possible ripple effects for cloud pricing and server design.