AMD Pitches Flash-as-DRAM Software to Cut Server Memory Costs by Up to 83 Percent

AMD this week started selling Mext, a software layer for its Epyc server line that swaps costly DRAM for NVMe flash storage and uses a machine learning model to hide the speed difference. The pitch is blunt: buy less memory, pay for a Mext license, and let the software figure out which pages actually need to live in fast RAM.

The architecture decision here is the story. Flash is orders of magnitude slower than DRAM, so naively tiering cold pages to an SSD would crater application performance. The Mext approach doesn't try to make flash faster. Instead, as reported by The Register, it predicts which memory pages an application will need before they're requested, pulling them back into DRAM ahead of time. The operating system sees the full memory footprint as installed DRAM. It doesn't know anything has been offloaded.

AMD acquired Mext in June. Gary Smerdon, the company's founder and now AMD's corporate vice president of business development, claims the technology could cut enterprise server costs by as much as 83 percent by replacing RDIMMs with flash. That number comes from the extreme end of the configuration range. According to The Register's coverage, the realistic headline figure is a 50/50 split: half DRAM, half flash.

The performance numbers depend heavily on the workload. Tom's Hardware reported that AMD's own tests against Synopsys' PrimeTime EDA suite showed 95 percent of full-DRAM performance at the 50/50 split, dropping to 80 percent when only 25 percent of the memory pool stayed as DRAM. Streaming workloads like Apache Spark do better: Notebookcheck noted AMD's claim of 99 percent performance retention with half the DRAM replaced. For multi-agent AI runs, AMD's data, as cited by BigGo Finance, shows a 90 percent improvement in performance-per-dollar at the 50/50 configuration.

The product isn't vapor. Notebookcheck reported the technology is already running on 145 systems, with AMD estimating around $5 million in customer savings. AMD has also confirmed it's running Mext inside its own IT infrastructure. Plans call for expanding to 2,000 servers in 2027, according to Tom's Hardware.

The context that makes this relevant right now: DRAM prices have been punishing. Wikipedia's DRAM entry notes prices have seen compounded increases exceeding 200 percent since early 2025, driven partly by HBM crowding out commodity capacity for AI accelerators. If you're building out an Epyc-based cluster and staring at a memory line item that's ballooned, Mext is at least worth modeling.

There are real caveats. The 83 percent cost savings figure assumes you're replacing the most expensive RDIMM configurations with the cheapest flash. Workloads that are genuinely memory-bandwidth-bound won't see the same results as EDA or Spark. And this is a licensed software product bundled with new Epyc hardware, so the cost math needs to include the Mext license itself, which AMD hasn't publicly priced.

What AMD is actually selling is the prediction model's accuracy, not the tiering concept itself. Memory tiering to flash isn't new. The question is whether the prefetch model is good enough to stay ahead of application demand. AMD says yes. Customers running latency-sensitive workloads will want to run their own benchmarks before committing to a thinner DRAM footprint.

Sources cited:
- The Register (https://www.theregister.com/systems/2026/11/11/amd-where-were-going-we-dont-need-as-much-dram/5308953)
- Tom's Hardware (https://www.tomshardware.com/pc-components/cpus/amd-shares-how-it-can-retain-95-percent-of-server-performance-with-just-half-the-dram-sidestepping-memory-shortages-with-mext-tech-powers-over-100-servers-already-running-with-plans-to-expand-to-2-000-servers-in-2027)
- Notebookcheck (https://www.notebookcheck.net/AMD-s-new-trick-lets-servers-use-half-as-much-DRAM-with-minimal-performance-loss.1448188.0.html)
- BigGo Finance (https://finance.biggo.com/news/f9e406ab-2eff-4853-a940-8455b2b9cb6f)