Recently, AMD officially announced an innovative memory solution called MEXT. This solution deeply integrates NVMe SSDs with physical DRAM through a software-defined approach, aiming to alleviate the severe memory cost and supply crisis in the current server market.
According to industry data provided by AMD, memory costs currently account for 75% of total server costs, and memory prices have recently surged from approximately $8 per GB to over $10 (about 72.4 RMB).
More seriously, data centers suffer from significant resource waste. Surveys by Microsoft Azure and Meta show that about 50% of memory is over-provisioned, while 50% to 80% of memory pages remain in a cold data state and are never effectively utilized.
The core of MEXT technology lies in using AI algorithms to offset the latency disadvantage of flash memory. Although the procurement cost of flash memory is only half that of DRAM, and its power consumption is reduced by 30 times, its access speed is 500 times slower than DRAM.
To this end, AMD integrated a lightweight AI model engine based on the Transformer architecture into the MEXT software to predict data needs in real time.
This engine occupies only one EPYC processor core, requires no GPU involvement, and can complete real-time training in less than 1 second, pushing pages about to be used from flash memory to physical memory in advance.
In actual tests, MEXT demonstrated remarkable efficiency. In Synopsys PrimeTime tests in the EDA field, even with a 50% reduction in physical memory usage, the system retained 95% of its original performance, improving performance per unit memory cost by 1.9 times.
Under more extreme testing with a 75% memory reduction, the system still maintained 80% of its performance level, achieving a 3.2-fold increase in performance per unit cost.
Currently, MEXT technology has entered its first phase of deployment, primarily targeting AMD EPYC server platforms equipped with NVMe drive bays. AMD stated that this solution has already avoided approximately $4.9 million (about 35.46 million RMB) in memory procurement expenses across the initial 145 deployed servers.
