Executive Summary: Cutting Server RAM Bills With Smart Flash Swap
- Half the DRAM, Near-Full Speed: AMD says Mext software lets Epyc boxes swap half their DRAM for NVMe flash while keeping 99% Spark speed and 95% chip-tool speed.
- No GPU Needed: Mext trains a tiny page predictor on one Epyc CPU core in under a second to pull cold pages back before apps ask.
- Early Cash Saved: Bought in June 2026, Mext runs on 145 early systems and claims roughly $5 million saved on skipped RAM sticks.
- Real Trade-Offs: Cutting DRAM down to 25% slows chip tools to 80%, and virtual setups burn one CPU core per VM.
Buying server RAM has turned into a brutal budget crunch for cloud teams. AI chip builds soaked up memory supply, pushing DRAM to as much as 90% of a big server’s total bill.
On November 11, 2026, The Register dug into AMD’s software pitch called Mext. Bought in June 2026 alongside AMD’s 6th-gen Epyc launch, Mext claims servers can drop half their DRAM for NVMe flash while keeping up to 99% speed.
Why Memory Eats Up 80% of Server Bills
Former Mext CEO Gary Smerdon, now a vice president at AMD, pointed to a huge waste inside data centers. Memory sticks make up 60% to 90% of a large server’s price tag. Yet most of that pricey RAM sits idle.
Smerdon cited a Meta study showing 50% to 80% of memory pages go untouched for five minutes or more, even on boxes marked 95% full. To a fast Epyc CPU, five idle minutes feels like years.
Flash storage costs 50 times less per bit and draws 30 times less power than DRAM. The catch is plain: NVMe flash runs 500 times slower, which wrecks app speed whenever Linux swaps to disk.
Smart Flash Swap Trained in Under One Second
Mext dodges disk swap lag by guessing which cold pages an app will touch next and loading them early. Its small model acts like text prediction, except it predicts upcoming RAM pages instead of words.
Pre-built models fail because every database touches memory in unique ways. So Mext uses one Epyc CPU core to train on live app habits in under one second, with zero GPU needed.
As long as pages land back in DRAM before the chip calls, Linux and apps never notice flash sat underneath. Buyers can order one terabyte of RAM instead of two and let cheap flash handle the rest.
AMD Speed Claims vs. VM Core Costs
Outside testers have not audited Mext yet, so IT teams should check AMD’s fine print closely. Across 145 early deployments, Smerdon estimates firms saved $5 million in skipped RAM buys, up to 83% off high-RAM server costs.
Still, Smerdon’s claim of zero hit softens when flash takes over. In AMD tests on Synopsys PrimeTime chip software, dropping half the DRAM held 95% speed. Dropping three-quarters of DRAM cut speed down to 80%.
Apache Spark fared better in AMD runs, holding 99% of base speed on half the RAM. Virtualized cloud hosts face one big catch too: Mext reserves one CPU core per VM or VPC, raising costs versus bare-metal containers.
Mext DRAM vs. Flash Speed Benchmarks
| Test Workload | DRAM vs. Flash Split | AMD Claimed Speed |
|---|---|---|
| Apache Spark Jobs | 50% DRAM / 50% NVMe Flash | ~99% base speed (vendor test) |
| Synopsys PrimeTime | 50% DRAM / 50% NVMe Flash | ~95% base speed (vendor test) |
| PrimeTime Heavy Swap | 25% DRAM / 75% NVMe Flash | ~80% base speed (20% drop) |
| Linux Setup | Kernel driver + RAM View app | AMD-only OEM software license |
What Server Buyers Say About Hardware Lock-In
Builders on Reddit hardware boards cheered relief from high RAM quotes, though sysadmins flagged vendor lock-in fast. Mext ships solely for AMD CPUs through server OEM partners like HPE.
Cutting DRAM orders hits during fierce memory supply squeezes affecting chip investors in Micron and Broadcom as well as debt-heavy clouds like CoreWeave.
Custom chip giants like Google with its latest TPUs fight memory limits daily too. AMD pairs Mext with a Linux driver and a RAM View tool so shops can spot safe servers before paying license fees.
Frequently Asked Questions
How does AMD’s Mext cut RAM needs without slowing Linux?
Mext runs a small model on one Epyc CPU core to guess upcoming memory page calls, pulling cold data from flash before apps ask.
Does Mext need GPUs to run memory predictions?
No. Model training finishes on a single Epyc CPU core in under one second, though virtual setups reserve one core per VM.
Can Intel servers run AMD’s Mext memory software?
No. AMD plans to license Mext only for AMD Epyc servers through OEM partners such as HPE.
