AMD proposes using NVMe alongside DRAM in AI servers

AMD described MEXT (MEXT Predictive Memory) software, which uses a transformer model trained directly on the server to proactively move 'cold' memory pages between DRAM and NVMe flash in Epyc servers, maintaining 99% of Apache Spark performance when half of the DRAM is replaced, according to vendor measurements.

Compiled by Sergey KostenchukPublished 2026-11-12Updated 2026-11-13

2026-11-12 Coding

💻 AMD replaces part of Epyc server DRAM with NVMe flash

AMD described MEXT (MEXT Predictive Memory) software, acquired from startup Mext: 'cold' memory pages are moved to NVMe disks, while a transformer that trains without a GPU proactively returns them to DRAM. According to AMD measurements, Apache Spark retains 99% of its performance when half of the DRAM is replaced. Flash is approximately 50 times cheaper than DRAM per bit, but approximately 500 times slower.

🌍 Memory accounts for 60–90% of the cost of large servers, and Bloomberg forecasts a DRAM shortage through 2030. MEXT reduces the cost of EDA, databases, analytics, and agentic AI pipelines on AMD hardware only.

👤 The model predicts memory pages, not tokens, and trains directly on the server. If it takes hold, a server with half the DRAM will become significantly cheaper with almost no loss in speed.

Source 1: https://www.theregister.com/systems/2026/11/11/amd-where-were-going-we-dont-need-as-much-dram/5308953 Source 2: https://www.amd.com/en/products/software/mext.html

Sources