AMD shares how it can retain 95% of server performance with just half the DRAM

AMD is promoting MEXT, a Linux kernel technology designed to reduce server DRAM needs by moving infrequently used memory pages to cheaper flash storage. Because flash is far slower than DRAM, MEXT uses an AI model to predict which pages applications will need and restore them before they are requested. The model reportedly learns each workload’s access patterns in under a second, runs on one EPYC processor core without a GPU, and continues training as the workload runs. AMD’s RAM View tool identifies frequently and rarely used pages, helping assess where memory reductions may be viable.

AMD reports that MEXT can preserve 95% of performance on Synopsys PrimeTime with half the DRAM, and 80% with only a quarter. Its other internal results include 99% performance at half capacity for a Spark SQL benchmark and for several concurrent AI agents; results vary by workload. The company has deployed MEXT on 145 servers and plans to reach 2,000 in 2027, while working with HPE on “MEXT Ready” EPYC systems. Customer licensing has no announced start date. The reported gains could ease costs amid memory shortages, but the tests are AMD’s own, their detailed methodology has not been disclosed, and MEXT’s usefulness depends on workload. Its cautious, limited rollout reflects the need to validate performance before broad adoption.