AMD Pitches Mext Software to Cut Server DRAM by Half Without Major Performance Loss
AMD's newly acquired Mext technology uses an AI prediction model to offload cold memory pages to flash, claiming up to 99% of DRAM performance at half the memory cost on certain workloads.
AMD is making a pointed pitch to data center buyers: cut your DRAM spend dramatically, keep almost all your performance. The argument rests on Mext, a startup AMD acquired in June 2026 whose software sits between a server's operating system and its physical memory, transparently tiering cold memory pages to NVMe flash while predicting which pages to pull back before applications need them.
The prediction layer is the whole trick. Flash is roughly 500 times slower than DRAM, so naive tiering would crater throughput. According to coverage in The Register, the Mext team uses an AI model to prefetch pages before they're demanded, making the latency gap less visible to the workload. Without that prefetch step, the numbers wouldn't hold up.
AMD's own IT infrastructure has already deployed the technology. As reported by Tom's Hardware, AMD is running more than 145 servers at 50% DRAM capacity using Mext. On Synopsys PrimeTime EDA workloads, the company claims those systems reach 95% of full-DRAM performance. For Apache Spark, the figures are more favorable: AMD says Mext retains 99% of performance with half the DRAM replaced by flash. Push the ratio further, to 25% DRAM and 75% flash, and performance drops to 80% of the DRAM-only baseline, per AMD's tests.
The cost framing comes from Gary Smerdon, Mext's former CEO and now AMD's corporate vice president of business development. In a statement covered by Crypto Briefing, Smerdon puts potential enterprise server cost reductions at up to 83%, driven primarily by swapping costly registered DIMMs for NVMe flash, which he says runs about 50 times cheaper per bit and draws 30 times less power. AMD's tally across its current 145-server deployment puts avoided DRAM costs at roughly $5 million, according to Notebookcheck.
The plan from here is scale. Tom's Hardware reports AMD is targeting 2,000 servers on Mext by 2027. A partnership with HPE is in phase one rollout, and Microsoft Azure and Oracle Cloud Infrastructure are listed as cloud hosts for the expanded deployment.
There are real caveats to hold onto. Every benchmark AMD is citing comes from AMD's own tests on AMD's own Epyc-based hardware. The two workloads highlighted, PrimeTime and Spark, both have memory access patterns that happen to suit tiering: they hit large datasets but tolerate some latency. The question is what Mext looks like on latency-sensitive inference workloads, which is precisely where data centers are spending the most on DRAM right now. AMD hasn't published those numbers yet, and Smerdon's 83% cost reduction figure assumes a configuration that takes the biggest acceptable performance hit.
The deeper structural point is that DRAM scarcity is a real constraint. The AI buildout has tightened the DRAM supply chain, and large model serving increasingly needs memory capacity that outpaces what DRAM alone can economically provide. That's the market condition AMD is selling into. Whether Mext's AI prefetch model is robust enough across the full range of enterprise and AI workloads, or whether it's tuned for the demos, is a question that gets answered when independent benchmarks show up. The 145-server baseline AMD is citing is a start, but it's also a number AMD controls entirely.
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)
- Crypto Briefing (https://cryptobriefing.com/amd-mext-software-cut-server-dram-costs/)
- Notebookcheck (https://www.notebookcheck.net/AMD-s-new-trick-lets-servers-use-half-as-much-DRAM-with-minimal-performance-loss.1448188.0.html)
This release was originally distributed via ETL Newswire. Visit The Register for the full story, related releases, and contact information.
Visit The Register →