AMD says its MEXT engine lets servers run on half the DRAM with little performance loss

If the claims hold beyond AMD's own tests, memory-constrained servers could need far less installed DRAM, though the rollout timing is still unclear.

AMD says its MEXT memory engine can let servers run with far less installed DRAM by moving rarely touched memory pages to flash storage and using a transformer-based model to fetch pages before the workload needs them. AMD claims that in its internal Synopsys PrimeTime workloads a server with half its DRAM reaches about 95% of full-server performance, and that the approach can cut physical DRAM needs by 20% to 75% depending on application and performance target. The technology runs as a Linux kernel module, trains itself within a second of deployment on a single EPYC core without a GPU, and keeps learning as the workload runs. AMD has deployed MEXT on 145 internal servers, plans to expand to 2,000 next year, and has a partnership with HPE to offer it to customers, though it has not said when the second rollout phase, which brings licenses and the RAM View tool, will begin. Because the claims come from AMD and its founder and the applications tested so far are narrow, the results remain largely unverified outside AMD's own environment.