AMD says an EPYC server can shed half its DRAM and still retain up to 99% of all-DRAM performance—if it’s running the right workload. Its Mext software moves less-active memory pages to NVMe SSD storage, then predicts when to bring them back. The pitch is less expensive memory capacity without leaving applications waiting on an SSD. SourceSource
Where the 99% comes from
The Register and El Chapuzas Informático report these AMD test results: SourceSource
These are vendor-reported results, not independent benchmark findings. Spark and PrimeTime also used different EPYC configurations. SourceSource
The SSD is still an SSD
Mext’s trick is prediction, not faster flash. According to AMD’s explanation, the software identifies infrequently accessed memory pages, moves them onto NVMe storage and anticipates when they’ll need to return to DRAM. If the pages arrive before the application asks, it avoids waiting for storage at that moment. SourceSource
AMD executive Gary Smerdon told The Register that Mext’s main model is a transformer derivative that predicts memory-page access rather than the next word. Training runs on an EPYC CPU, without requiring a GPU. The AI has a refreshingly specific assignment: have the data ready. Source
Mind the savings label—and the rollout
El Chapuzas Informático describes an AMD internal case involving 86 servers and roughly 58 TB of additional capacity. AMD estimated that adding that capacity with DRAM would have cost about $3.3 million, versus approximately $601,000 with Mext: a claimed 82% saving on the memory expansion. That is not an 82%-off coupon for the entire server. Source
The Register reports that AMD is initially working with OEMs such as HPE on “Mext ready” systems. A later phase would introduce licenses sold through OEM partners, exclusively for AMD hardware, and EPYC platforms with AMD’s RAM View workload-assessment tool preinstalled. Neither report gives a license price or a firm general-availability date. SourceSource
Linux integration uses a kernel driver, according to The Register. El Chapuzas Informático also reports a dedicated NVMe SSD, disabled conventional swap and two CPU cores, one for AI and one for its service. The Register describes one core for model training and one per VM or VPC in virtualized environments. The total CPU overhead for a given deployment remains unclear. SourceSource
Before the DRAM shopping list gets shorter, the workload needs a test run.
Comments