Johnathan Russell

AMD Taps Flash-Based MEXT to Cut DRAM Needs by Up to 75% $AMD’s MEXT software moves inactive memory pages to NVMe flash, then uses AI to predict future accesses and prefetch those pages into DRAM before applications request them. The reported target is a 20–75% reduction in physical DRAM requirements on EPYC servers, depending on the workload. The technical advantage is hiding flash latency through predictive prefetching. DRAM serves active data while flash provides cheaper capacity. This means performance depends on getting pages back into DRAM before they are needed; unpredictable access patterns can limit the benefit. AMD says MEXT runs on Linux using two CPU cores, with no GPU required.