AMD has learned to save on expensive server memory: its MEXT technology transfers rarely used data from DRAM to much cheaper SSDs, and AI algorithms help prevent a noticeable drop in performance. On the first 145 server systems, the company estimated savings of approximately $5 million.
Image Source: AMD
How told According to The Register, AMD Vice President of Business Development Gary Smerdon, formerly head of MEXT, in large servers, RAM modules can account for 60 to 90% of the cost of the equipment. Replacing some DRAM with significantly cheaper flash memory can reduce costs without necessarily reducing the amount of memory available to applications.
The problem is the speed difference. According to Smerdon, flash memory can be 50 times cheaper than DRAM on a per-bit basis and consume 30 times less power, but be about 500 times slower. Regularly offloading data to an SSD can seriously slow down applications, especially latency-sensitive ones.
MEXT uses a machine learning model that monitors memory accesses and tries to determine which data pages the processor will need next. Inactive pages are moved to the drive and returned to DRAM before being accessed again. This process should remain invisible to the operating system and applications.
The prediction engine is based on a model derived from the Transformer architecture. Instead of the next word, as in language models, it predicts the next page of memory needed. According to Smerdon, initial training takes less than a second on a single EPYC processor core, after which the model continues to adapt to the nature of the load. A graphics accelerator is not required for its operation.
Results vary by application. In AMD’s testing, a system running Apache Spark retained 99% of its original performance when replacing half of the DRAM with flash. In Synopsys PrimeTime, a similar configuration provided 95% performance, and when physical DRAM was reduced to a quarter of the original size, the figure dropped to 80%.
AMD also published results from its internal implementation of MEXT. Across the company’s 86 servers, the software provided approximately 58 TB of additional available memory. Purchasing a comparable amount of DRAM would cost approximately $3.3 million, while the cost of implementing MEXT was approximately $601 thousand. Thus, the estimated savings reached $2.78 million, or 82% of the cost of expanding memory using the traditional method.
The software is intended primarily for servers based on AMD EPYC processors. On Linux systems, a kernel module is provided for its operation, and the RAM View utility is offered for load analysis. AMD is working with server manufacturers, including HPE, to prepare hardware for MEXT. Commercial distribution of licenses through partners is planned for the next stage of implementation.
AMD acquired developer MEXT in June 2026. The technology’s capabilities were also presented along with EPYC Venice server processors based on Zen 6 architecture.
If you notice an error, select it with the mouse and press CTRL+ENTER.
