AMD has turned its June acquisition of MEXT into a product. The software makes NVMe flash look like extra DRAM to Linux and its applications, and AMD says it can cut the DRAM an EPYC server needs by 20% to 75% while keeping performance close to an all-DRAM machine. Wccftech's write-up of AMD's launch slides, published November 10, also shows a Hewlett Packard Enterprise partnership around "MEXT Ready" on-prem servers.
Why it matters: memory has become the expensive part of a server. AMD's slides, as reported by Wccftech, put DRAM at up to 75% of server cost, and say big operators overprovision it by about 50% while much of what's installed sits idle holding cold pages. If software can safely move those cold pages to flash, which AMD's slides describe as cheaper and far less power-hungry than DRAM, buyers can order smaller DIMM configurations in the middle of a memory shortage. BigGo's analysis points out the flip side for memory makers: if overprovisioning falls, longer-term server DRAM demand could be revised down.
How it works
According to AMD's product page and case study, MEXT moves infrequently used memory pages out to a dedicated NVMe drive. An AI engine predicts which pages an application is about to touch and pulls them back into DRAM before they're requested, so the application still sees its data in memory.
Wccftech lists three pieces: a lightweight Linux kernel driver, the prediction engine (an ensemble of models that AMD says trains in under a second and runs on a single EPYC core with no GPU), and RAM View, a dashboard showing real memory use. BigGo reports that MEXT itself is a paid license while RAM View is free.
AMD's stated requirements are modest but specific:
| Requirement | What AMD says |
|---|---|
| Workload | Memory-constrained, ideally needing more than 64 GB of DRAM |
| Storage | A dedicated NVMe drive; swap must be disabled |
| CPU | Two cores (one for the AI engine, one for the daemon); no GPU |
| OS | Most common distributions, incl. RHEL, Rocky, AlmaLinux, Oracle Linux, SUSE, Debian, Ubuntu |
| How to get it | Contact AMD sales for a proof of concept; no public price or download |
BigGo adds two practical notes: each guest VM needs a core for prediction, and AMD recommends provisioning roughly twice the target expansion capacity in flash to limit SSD wear.
AMD's performance claims
All of these are AMD's own benchmarks, reported by Wccftech and BigGo, and not independently tested:
| Workload | Configuration | AMD's claimed result |
|---|---|---|
| Synopsys PrimeTime (chip design) | DRAM cut 50% | Up to 95% of performance, ~1.9x perf per dollar |
| Synopsys PrimeTime | DRAM cut 75% | ~80% of performance, ~3.2x perf per dollar |
| Spark TPC-DS on Azure (5th Gen EPYC) | 50% DRAM / 50% MEXT | Up to 99% of baseline |
| Five-agent AI workflow | 50% DRAM / 50% MEXT | +90% perf per dollar |
| Memory-hungry agents | 50% DRAM / 50% MEXT | +70% perf per dollar |
BigGo reports that AMD's Gary Smerdon acknowledged some workloads haven't met expectations and that customers need to verify long-term benefits themselves. That's the right frame: the gains depend on how much of a workload's memory is actually cold.
AMD's own deployment, and two sets of numbers
AMD's IT department is the reference customer, but its figures differ depending on the document:
| Case study (AMD, Oct 30, updated Nov 3) | Launch materials (per Wccftech, BigGo) | |
|---|---|---|
| Initial servers | ~86 production servers | 145 servers |
| DRAM cost avoided | ~$2.78M of ~$3.3M (82%) | ~$4.9M (83%) |
| Scale target | ~700 servers by March 2027; 2,000 by end of 2027 | 2,000 servers in 2027 |
The case study says the 86-server rollout added about 58 TB of usable memory (a 20% increase) for roughly $601,000, and projects more than $60 million in savings at 2,000 servers. The simplest reading is that the deployment grew between the case study and the launch, but AMD hasn't reconciled the figures publicly. The case study names 4th and 5th Gen EPYC servers; the launch benchmarks use 6th Gen.
Who should care
- Teams buying EPYC servers for EDA, analytics, databases or agent workloads: ask AMD or HPE for a MEXT proof of concept before you lock in DIMM counts. Measure your own tail latency and SSD wear, not just averages.
- Anyone on Intel or Arm: AMD's materials only describe EPYC systems, though it lists on-prem and the big clouds as deployment targets.
- Memory-market watchers: this is one more sign that buyers are engineering around DRAM prices. Nvidia's CMX platform, BigGo notes, takes a related flash-tier approach but targets AI-inference KV cache rather than general server memory.
What's missing: public pricing, a list of supported CPU models, and any third-party benchmarks. Until those appear, treat the 95% figure as a best case for the right workload, not a default.
Sources
- AMD — AMD IT Avoids 82% of DRAM Expansion Costs with AMD MEXT Software (case study) amd.com
- AMD — MEXT: Making Flash Work Like DRAM (product page) amd.com
- Wccftech — AMD Unveils MEXT Memory Solution wccftech.com
- BigGo Finance — AMD Launches MEXT Predictive Memory Software, Cutting DRAM Requirements by Up to 75% finance.biggo.com