AMD expands MEXT memory-offload tech, cutting DRAM needs by up to 75%
AMD acquired MEXT, a predictive memory engine that offloads cold DRAM pages to flash storage and uses an AI model to prefetch data before it's needed. AMD says internal tests, such as with Synopsys PrimeTime, show 95% of full-DRAM performance using only 50% of physical DRAM, and the company plans to scale deployment from 145 internal servers to 2,000 next year, alongside a rollout partnership with HPE.