Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

AMD promotes MEXT predictive memory software

AMD presented MEXT on November 10, using predictive software and flash to reduce server DRAM needs in selected workloads.

Published 16 Nov 2026· 5 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

AMD promoted its MEXT predictive memory software at the EPYC Venice event in Las Vegas on November 10, 2026, presenting flash storage as an additional memory tier for servers. For research computing teams, the development offers a potential way to stretch DRAM capacity, but the performance evidence comes from AMD’s chosen workloads and access remains through customer channels. Phoronix reported the presentation.

What changed in November

The November presentation follows AMD’s June 15 announcement that it was acquiring MEXT. In that announcement, AMD described predictive memory as a way to expand usable capacity for AI and general computing, and said it expected to integrate the technology across its data center portfolio. The new development is the software’s promotion with EPYC and the publication of detailed performance evidence, rather than another acquisition announcement.

TrendForce’s November 11 coverage places the rollout against memory supply constraints and reports AMD’s push toward enterprise commercialization. Its article explicitly draws on Wccftech and Commercial Times. Those reports provide coverage of the rollout, rather than independent reproduction of AMD’s benchmark results.

How predictive memory uses flash

AMD’s product description says MEXT moves cold memory pages to flash, predicts which pages will soon be needed and brings them back to DRAM ahead of application requests. The aim is to keep the pages an application needs ready in DRAM while using flash for less active data. This is software managing a memory hierarchy; the announcement does not introduce a new DRAM chip or establish a replacement for GPU HBM.

The distinction matters when reading the performance claims. MEXT’s proposed benefit depends on anticipating page reuse, rather than making every flash access physically equivalent to a DRAM access. AMD’s June announcement framed the technology as expanding usable memory capacity while maintaining performance and efficiency. A research workload’s access pattern therefore belongs in any evaluation of that proposition, alongside its total memory requirement.

What AMD’s agentic AI tests show

AMD’s November agentic AI white paper reports internal testing conducted on October 20, 2026. Its reference system used two EPYC 9996 processors, Ubuntu 24.04.5 LTS, Linux 6.17.12 and MEXT 1.2, with a Solidigm 3.84 TB NVMe drive. The baseline DRAM allocation was 1,024.7 GiB per socket.

AMD-reported performance relative to its all-DRAM baseline
Retained DRAM Added MEXT memory pool Five-agent mixed phase Solo image-processing agent
50% 512 GB Near baseline 89.7%
25% 768 GB 85.3% 88.4%

The mixed-phase metric is a geometric mean across five agent personas at a 640-worker step. The separate image-processing test used the mtpillow persona at 2,560 workers. AMD averaged three runs per scenario. These are workload-specific throughput results: retaining half the DRAM did not preserve the same proportion of performance in both tests. AMD also says results vary with workload, software, storage and system configuration.

The research implication is to compare a proposed memory reduction against an acceptable throughput loss. A centre can use performance profiles for benchmarking to think about comparisons across workloads, while keeping AMD’s particular agent suite separate from acceptance tests for local research codes.

Linux support and access have practical conditions

As described in AMD’s official FAQ on November 16, 2026, deployment requires a dedicated NVMe drive and disabled swap. The software uses two CPU cores, one for its AI engine and one for its daemon, and requires no GPU. AMD lists RHEL, Rocky, AlmaLinux, Oracle Linux, SUSE Linux, Debian and Ubuntu among supported distributions, with environment confirmation through AMD sales.

This is narrower than a guarantee that every Linux installation will work. Phoronix’s Michael Larabel reported on November 10 that AMD offered a Linux driver but that MEXT was proprietary and available through customer or sales channels, without a public, freely available release. AMD’s current FAQ directs prospective users to a sales representative to begin a proof of concept. That is the documented access route for an institution considering an evaluation.

What research computing teams can take from the rollout

For universities, the immediate decision is whether to evaluate MEXT on a suitable workload. AMD identifies memory-intensive applications needing more than 64 GB of DRAM as its strongest candidates, and lists agentic AI, databases and analytics, electronic design automation, and media workloads. That scope gives research computing offices a starting point for selecting trial jobs; it does not establish performance for every application in a shared cluster.

As an evaluation approach, a team could compare the same research job with MEXT disabled and enabled, documenting retained DRAM, flash allocation, concurrency and completion time. Keeping the input data and application version fixed would help distinguish a memory-tiering effect from a change elsewhere in the software environment. The record should identify the kernel, MEXT version and storage device so that a later run can be compared meaningfully. These are considerations for a reproducible AI experiment and for computational reproducibility.

For research offices preparing infrastructure proposals, the useful outcome would be a measured capacity-and-throughput trade-off for the intended project. The reported ability to reduce physical DRAM is a reason to test that trade-off, rather than evidence that an existing memory specification can simply be reduced by the headline percentage.

What happens next

TrendForce, citing Wccftech, reports an AMD internal rollout starting with 145 servers and a goal of reaching 2,000 in 2027. It also reports work with HPE on enterprise adoption. Those are reported deployment plans, rather than completed expansion or a dated public software release. For external research institutions, the next concrete step described by AMD remains a proof of concept arranged through its sales channel.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →