AMD Watch posted an update
AMD’s Mext software uses machine learning to predict which memory pages applications will need, moving less-used data from DRAM to flash and bringing it back when needed. Notebookcheck reports that AMD says this can let some Epyc servers use half as much DRAM with limited performance loss.
Why it mattersIn AMD’s tests, moving half the memory to flash retained 95% of performance in Synopsys PrimeTime and 99% in Apache Spark. Those are company-reported results, and the trade-off depends on the workload. In virtualised environments, Mext also needs one CPU core per virtual machine. AMD says the software is being prepared for enterprise customers, but pricing and wider availability dates have not been disclosed. Cheaper memory is appealing; the benchmark details are where the small print earns its keep.
Discuss: Would you trade some performance for lower server-memory costs, or is predictable speed more valuable for your workloads?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.