AMD says Mext can expand server memory with less DRAM

AMD says its Mext Predictive Memory software can expand usable server memory with NVMe flash while reducing the amount of physical DRAM required. The company estimates that one internal deployment added about 58 terabytes of usable capacity for roughly $601,000, compared with an estimated $3.3 million for an equivalent DRAM expansion. The comparison is projected, not a record of savings already realized.

Mext uses AI models to predict which memory pages an application will need and moves those pages from flash into DRAM ahead of a request. Less frequently used pages are stored on flash, which is slower than DRAM but less expensive. The approach aims to keep workloads from waiting for data to be retrieved after memory pressure develops, as can happen with conventional swap or page-file systems.

AMD says certain workloads retain 95% to 99% of full-DRAM performance with less physical memory. In one cited result, its internal testing of Synopsys PrimeTime showed 95% of performance with half the DRAM. AMD also reported that mixed DRAM-and-Mext configurations delivered 1.9 to 3.5 times the performance per dollar. These are company-reported results; the material does not provide...