Latest

Technology

AMD MEXT turns your NVMe into DRAM with AI, and it is not open source

AMD MEXT uses AI to move cold pages from DRAM to NVMe and present them as memory. A module for Linux, 1.74x per dollar of DRAM and a closed license.

AMD MEXT turns your NVMe into DRAM with AI, and it is not open source
AMD MEXT turns your NVMe into DRAM with AI, and it is not open source

AMD unveiled MEXT at the EPYC Venice event held in Las Vegas, and it did so with the most pragmatic argument available in 2026: memory is extremely expensive and half of what is bought goes unused. The software, acquired by AMD this summer, moves cold pages from RAM to an NVMe SSD and brings them back before the application requests them, with a prediction engine that decides the order. The result, as described in the manufacturer's documentation, is a server that can run on half the physical DRAM and perform about the same.

The proposal fits the market moment. TrendForce puts the quarterly rise in DRAM at 20%, and AI is monopolizing manufacturing capacity. The price of RAM has soared to the point where memory has become the most uncomfortable line on the bill, and buying less DRAM while making the most of an already-installed SSD sounds like an obvious solution.

The uncomfortable detail comes on the final slides. MEXT runs on Linux, is presented as a kernel module that installs in minutes and promises compatibility with any distribution, but it is not open source. It is distributed through AMD's commercial channels, under license, and there is no public version or repository. It is the exception in the company's strategy, which has spent a decade publishing its drivers as open source, and it deserves more than a footnote.

How does it work, in three steps?

The system rests on three pieces that arrive in a single package. The first is a loadable kernel module that does not modify the standard kernel and is compiled at install time to match the distribution and system version. It is responsible for collecting telemetry on processes and memory pages. The second is the AI engine, a user-space process that runs on a single CPU core and does not need a GPU: it predicts which pages will be used again and in what order.

The third piece is the daemon, which carries out the actual movement of pages between DRAM and flash. The full cycle begins with an eviction: the driver detects pages that are no longer in use and moves them to the NVMe. Then a set of models, from lightweight heuristics to neural networks, predicts which of those pages will be needed soon and returns them to RAM before the application requests them. The adjustment is dynamic, and the documentation itself explains that the system learns from its hits and misses.

The analogy with language models is not gratuitous, and AMD uses it: instead of predicting the next word, the engine predicts the next memory access. The MEXT Predictive Memory technical sheet details the requirements, which include eight physical cores or vCPUs, at least 64 GB of DRAM and an unmounted NVMe SSD presented to the system as a native device. The default space dedicated to MEXT memory is 20% of physical DRAM, adjustable with a command.

The underlying idea is not new. Operating systems have spent decades moving pages between memory and disk with simple age-based algorithms, and what changes here is the predictor that decides what to bring back and when.

There is a hardware limitation worth stressing because it does not appear in the headlines. The SSD used as extended memory needs a matching write endurance: the manufacturer's table calls for a consumer drive for workstations and an enterprise-class one for servers, with a minimum of one full drive write per day in the first case. Using a cheap disk for this is a quick way to destroy it.

The numbers and the context in which they were measured

AMD publishes a timing analysis case with Synopsys PrimeTime on an EPYC 9755 with 128 cores. The reference configuration carries 1.5 TB of DRAM; the configuration with MEXT carries 768 GB of RAM and another 768 GB of flash memory from a 3.84 TB Samsung drive. The manufacturer's internal measurement puts the second configuration at around 1.74 times better performance per dollar of DRAM. The same documentation warns that this metric does not include the cost of the NVMe, the MEXT license or the total cost of ownership.

That nuance matters more than it seems. The real saving depends on the price of the SSD, the cost of the license, which is not public, and the type of workload. A database server with a dataset that does not fit in RAM and an unpredictable access pattern does not behave like a timing analysis tool with well-defined cold pages. Prediction works best when there are patterns, and not all workloads have them.

Where the argument holds up most strongly is in hardware that has already been bought. A server with 1 TB of RAM installed and applications that use 400 GB is paying for capacity it does not use. There MEXT does not replace memory; it makes it pay off.

In a data center with hundreds of nodes, that capacity margin is hard cash, and it explains why AMD is pushing it at the peak of memory prices.

The problem of the closed license

The surprise is not in the technology, but in the distribution model. AMD has built its reputation in the Linux ecosystem by publishing drivers, firmware and tools as open source, and has used that argument against competitors that ship blobs. MEXT breaks that pattern: there is support for Linux, yes, and the company promotes it as a module that works with any system, but the code cannot be read, audited or modified.

For a systems administrator, that raises three reasonable questions. The first is what happens if the module fails on a new kernel, something common in infrastructure that updates late. The second is how to audit a component that touches the memory management of the entire server and runs with kernel privileges. The third is what happens to the license if the customer changes server vendor or decides to migrate to the cloud.

None of the three has a public answer. AMD says MEXT is available through its commercial channels and that it could be opened up in the future, a formula that in the industry usually means the decision depends on competitive pressure. The problem is that tiered memory is a space where open alternatives already exist, some integrated into the Linux kernel itself, without a prediction engine but with the advantage of being able to look inside them.

The honest comparison is not MEXT against nothing, it is MEXT against a kernel with poorly configured tiered memory, which is what most servers have today. That is the product's real rival, and it does not appear on any slide.

There is also a dependency risk that goes beyond the code. If the saving on DRAM becomes the reason to buy AMD servers, memory management software turns into one more commercial lever. That is not illegitimate, but it changes the conversation: it stops being a technical improvement the customer evaluates and becomes a recurring fee the customer pays.

Memory as a product

AMD's move is consistent with the market and at the same time reveals a tension that will recur. For years, the software that made hardware perform better was free and open, because it sold processors. When the bottleneck shifts to memory and each gigabyte costs what it costs, the software that stretches that gigabyte acquires value of its own and the manufacturer starts charging for it. The crisis is already visible in consumer products: the next generation of consoles is being born conditioned by the price of memory.

It remains to be seen whether customers accept the trade-off. An infrastructure director who can halve the DRAM in an order of 500 servers has a powerful argument for paying licenses, even if the code is closed. Whoever manages a cluster with custom kernels and internal audits will have to decide whether that saving is worth the black box. Both answers are legitimate and neither is settled by reading the press release.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading