AMD is turning to software and SSD storage to reduce the cost of expanding server memory. Its MEXT technology moves less frequently used data out of DRAM and predicts what applications will need next, bringing it back before it is requested.
A November case study describes an initial deployment on 86 of the company’s production servers. It added about 58 TB of usable capacity. AMD estimates that the deployment cost about $601,000, against roughly $3.3 million for equivalent additional DRAM capacity.
Storage is not suddenly as fast as memory
The approach relies on giving different kinds of hardware different jobs. DRAM serves data that applications need immediately; NVMe flash provides a less expensive pool for data that can be moved out of the fast tier.
Prediction is intended to limit the performance penalty from that movement. It does not eliminate the physical difference between flash and DRAM, and its effectiveness depends on how a workload accesses data.
Savings depend on the comparison
The reported 82% saving concerns the cost of the capacity expansion in AMD’s internal example. It is not an 82% reduction in every server’s total cost, nor an independently established result for every application.
A separate November white paper examines performance per dollar spent on DRAM for agentic-AI workloads. Such comparisons need to be read alongside the tested configuration and software, rather than treated as a promise for any existing machine.
AMD acquired MEXT in June and began the internal rollout in September. The current development is evidence from that deployment and the November technical material, not a new acquisition. As memory demand grows, the commercial question is whether software can stretch existing equipment economically without introducing an unacceptable delay for the applications running on it.






