AMD's new trick lets servers use half as much DRAM with minimal performance loss - Notebookcheck
AMD has developed Mext software that uses AI to move rarely accessed data from DRAM to cheaper SSDs, potentially halving server memory costs while maintaining near-original performance levels.
The technology is already running on 145 systems, with AMD estimating around $5 million in customer savings. By using machine learning to predict which memory pages applications will need next, Mext moves less frequently used data from expensive DRAM to much cheaper SSDs.
Memory Cost Reduction
According to Gary Smerdon, Mext's former CEO and now AMD's corporate vice president of business development, memory can account for 60% to 90% of the cost of large servers. Replacing some of that DRAM with SSD storage could therefore save customers a considerable amount of money.
Memory can account for 60 % to 90 % of the cost of large servers.
Performance Impact
The approach is similar to swap files or the Windows page file, which move data from RAM to storage when memory is running low. However, flash storage can be up to 500 times slower than DRAM, so this usually comes with a performance penalty.
Mext aims to avoid much of that slowdown by using machine learning to predict which memory pages applications will need and move them back into RAM before they are requested. Smerdon cited Meta research showing that 50% to 80% of memory pages had not been accessed for at least five minutes.
Implementation Results
AMD's tests show that Mext's results depend on the workload. In tests with Synopsys PrimeTime, Mext replaced half of the system's DRAM while retaining 95% of its original performance. With Apache Spark, it retained 99% of performance after half of the DRAM was replaced with flash.
The software uses several machine-learning algorithms, including a model derived from transformer technology, with initial training taking less than a second on a single Epyc CPU core. The model then adapts as workloads change without requiring a GPU.
Future Plans
AMD plans to license Mext through its OEM partners, and the software will work exclusively with AMD hardware. The company is working with manufacturers such as HPE to prepare compatible Epyc systems.
However, there are limitations. In virtualized environments, the software requires one CPU core per virtual machine or virtual private cloud instance, which could create significant overhead on servers running many virtual machines. AMD is also developing a RAM View application to help customers identify which servers and workloads could benefit from the technology.
Related
