AMD pitches Mext software that lets servers run on half the DRAM
AMD is selling software from its Mext acquisition that moves cold memory pages from DRAM to cheaper flash storage, a pitch aimed at buyers squeezed by the memory chip shortage.
Kiln, covering chips, data centers and power
, 2 min read
145Servers running Mext so far, according to AMD
What to know
- AMD's Mext software uses an AI model to predict which memory pages a server will need and keeps the rest on NVMe flash.
- In AMD's tests reported by The Register, half the DRAM gave 95% of full performance on one chip-design workload and 99% on Apache Spark.
- AMD says the first 145 systems avoided about $5 million in DRAM purchases; the software runs only on AMD EPYC servers.
AMD is pitching data center buyers on software that lets a server do most of its work with half the usual DRAM. The software, called Mext, came with AMD's purchase of the startup of the same name in June. It moves memory pages that are not in use out to cheaper NVMe flash storage and pulls them back before the operating system or application asks for them.
The details come from an interview AMD's Gary Smerdon, Mext's former chief executive and now a corporate vice president of business development at AMD, gave to The Register, published November 11. HostingJournalist and Tom's Hardware have since reported the same claims. AMD has not published independent benchmarks, and the performance figures below are AMD's own.
How it works
Mext works like a smarter version of swap space or a Windows page file. A small model derived from the transformer design behind chatbots watches which memory pages a program touches and predicts which it will need next. Pages that have gone cold sit on flash. Pages about to be used are moved back into DRAM ahead of time, so the software sees them as always in memory.
According to The Register, the model trains on a single core of an AMD EPYC processor, with no graphics processor, and the first training pass takes under a second. In virtual machines the software needs one CPU core per machine, a cost buyers have to set against the memory they save. On Linux it needs only a kernel driver.
What AMD claims
In AMD's tests on Synopsys PrimeTime, a chip-design tool that uses a lot of memory, putting half the memory on flash kept 95% of the performance of an all-DRAM server. Cutting DRAM to a quarter of the total dropped performance to 80%. On Apache Spark, a data processing system, AMD said half the DRAM delivered 99% of full performance. Smerdon said cold pages can move to flash with "literally no impact on performance"; The Register noted that its own figures show some penalty.
AMD says Mext is running on 145 systems so far, and that those customers avoided an estimated $5 million in DRAM purchases. Smerdon put the possible saving on server costs as high as 83%, because memory can make up most of the price of a large-memory machine. Tom's Hardware reported that AMD aims to expand the deployment to 2,000 servers next year.
The software runs only on AMD's EPYC server chips, including the new Venice generation. AMD plans to sell it as a license through server makers such as HPE, starting with memory-hungry jobs like chip design and content creation and moving later to databases and AI agent pipelines.
Why it matters
DRAM has become one of the scarcest and most expensive parts of a server, because memory makers have shifted production toward the high-bandwidth memory (HBM) that AI accelerators such as Google's latest TPUs use. If Mext performs as AMD says on real workloads, it gives buyers a way to cut how much DRAM they need, and gives AMD a selling point for EPYC that Intel and Arm-based server chips cannot match. The open question is how it holds up on latency-sensitive work, where a wrong guess about which page comes next is more costly.
Sources
- AMD: Where we're going we don't need (as much) DRAM, The Register, November 11, 2026
- AMD Says Mext Can Halve DRAM Needs on EPYC Servers, HostingJournalist, November 11, 2026
- This week in Tom's Hardware Premium: November 14, 2026, Tom's Hardware, November 14, 2026
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.