Key points
- MEXT uses AI-based predictions to keep frequently accessed data in DRAM and move less-used data to NVMe flash.
- AMD reports that systems combining DRAM and MEXT can retain up to 99% of the performance of a full-DRAM configuration in selected tests.
- The technology requires a Linux driver, runs on a single EPYC CPU core, and does not necessitate server hardware or software changes, according to AMD.
Context
MEXT is aimed at reducing data centers' exposure to expensive and scarce DRAM. Flash is cheaper and more energy-efficient but much slower, so AMD's prediction engine preloads data likely to be requested soon. The technology is already being deployed in AMD's own server infrastructure, with possible expansion into consumer PCs.
What this means for you
Directly applicable
You can evaluate MEXT as a memory-tiering option for EPYC servers, particularly where Linux workloads could benefit from lower DRAM demand without major application changes.
Who benefits: Developers and technical decision-makers operating memory-intensive EPYC servers can assess whether flash-backed memory extension fits their cost and performance constraints.
Share article: