AMD says its Mext Predictive Memory software can add server capacity at a projected cost far below an equivalent DRAM expansion, a comparison relevant as AI computing raises demand for fast memory.
The first deployment spanned 86 to 145 hosts; AMD puts the added usable memory at about 58 terabytes and the cost at roughly $601,000. For the same capacity, the company estimated a DRAM-based expansion would cost $3.3 million. The gap is about $2.7 million, or roughly 82% of the DRAM estimate. That is a projected comparison, not savings already recorded.
Mext Predictive Memory uses AI models to forecast which application data is likely to be needed and move it from NVMe flash storage into DRAM before an application requests it. DRAM is the faster working memory, while flash provides another storage tier. In conventional configurations, Linux swap or Windows page files can move data after memory pressure develops, leaving applications waiting while needed information returns. Mext is designed to stage likely-needed data ahead of a request.
AMD said certain workloads retained 95% to 99% of performance even after systems lost more than half of their physical DRAM. In a separate comparison, the company estimated mixed DRAM-and-Mext configurations produced 1.9 to 3.5 times the performance per dollar.
AI and high-performance computing workloads can require large pools of fast memory, while DRAM prices have been rising alongside demand. Those pressures add to the incentive to find ways to expand capacity.
Mext Predictive Memory launched in April 2026, and AMD acquired Mext on June 15, about two months later. As AMD extends the software, it has set a cumulative savings target above $60 million through the close of 2027.
