
AMD Presents MEXT Software to Expand Effective Server Memory
The AMW Read
AMD's presentation adds a memory-tiering approach to its infrastructure offering, but the reported benefits depend on workload and configuration.
AMD Presents MEXT Software to Expand Effective Server Memory
AMD presented MEXT, software that moves infrequently accessed memory pages from DRAM to SSD storage and predicts which pages to restore. The company acquired startup MEXT in June. The Linux-based system includes a driver, an AI prediction engine and RAM View, a tool for inspecting memory access patterns. AMD says the approach can expand effective memory capacity by up to three times without specialized memory hardware.
The technology targets a cost constraint in AI infrastructure: servers may need enough DRAM for peak demand even when much of that memory is accessed infrequently. AMD's presentation extends its earlier case for EPYC CPUs in agentic AI to include memory efficiency. MEXT could reduce the DRAM needed for suitable workloads, though the benefit depends on access patterns and the cost of retrieving data from SSDs. The reported examples do not establish a universal performance or savings rate.
Builders should test memory access patterns and application performance before reducing DRAM. In one reported Azure example, a three-node cluster retained 99% of its performance with half the memory; the article does not specify the full configuration. In another, Synopsys PrimeTime on Azure HX reached up to 1.4 times the throughput by running four concurrent jobs instead of two. HPE plans to offer MEXT-ready servers first and MEXT licenses in a later stage.


