AMD's AI trick squeezes 95% performance from half the RAM, promising huge server savings

AMD has unveiled a technology called MEXT that aims to drastically reduce the physical DRAM needed in servers while maintaining nearly full performance. The system works by offloading less frequently used 'cold' memory pages to cheaper flash storage, then uses an artificial intelligence model to predictively pull those pages back into active DRAM before they're needed by applications. AMD claims it can retain 95% of performance in some workloads using just 50% of the DRAM, with potential reductions ranging from 20% to 75% depending on the application.

This innovation arrives as a direct counter to soaring DRAM prices driven by AI industry demand, affecting both consumer and enterprise hardware costs. If MEXT delivers as promised, it could transform data center economics by allowing companies to build powerful servers with significantly less expensive memory. The technology is currently being tested internally on 145 AMD servers, with plans to expand to 2,000 next year and a partnership with HPE to bring 'MEXT Ready' servers to market. While still in early rollout phases with narrow applications, this AI-driven memory management represents a potentially game-changing approach to combating hardware shortages and cost inflation.