🧠AMD's Mext Turns Flash Into DRAM, Cutting Server Memory Costs Up to 83%

Your RAM bill might shrink by half this year

AMD just put a price tag on memory you don't need to buy. Its Mext technology uses AI to push rarely touched memory pages out to flash, then yank them back into DRAM before your code notices. Flash is 500x slower than DRAM, so the trick is predicting what's needed and when. If you run memory-hungry workloads like EDA, content creation, databases, or AI pipelines, this changes your hardware math. The software makes flash look like DRAM to the OS, and on Linux it needs only a kernel driver. A separate RAM View app helps you figure out which systems and workloads are worth enabling it on. The catch: it's exclusive to AMD hardware and sold as a license through OEM partners. The numbers are the real story. Offloading half of memory to flash still delivered 95% of DRAM-only performance in tests, and in some scenarios AMD sees 99%. Drop to 25% DRAM and 75% flash, and performance falls to 80% of baseline. So far Mext runs on 145 systems, avoiding $5 million in DRAM costs. The model trains on a single Epyc core in under a second, with no GPU needed. Right now AMD is working with HPE and others on "Mext ready" systems, with license sales and Epyc platforms shipping RAM View preinstalled in phase two.