Evolution of memory bandwidth per socket and per core
ServeTheHome has put together an interesting article for anyone looking to buy a server or interested in how memory bandwidth has changed between Intel and AMD recently. They trace his performance from 2013 to the present and track some theoretical results of AMD’s Bergamo in different ways to give a bird’s-eye view of how the two companies have changed over the years. It also shows why AMD’s EPYC has managed to capture a significant portion of the market once dominated by Intel’s Xeon.
First, looking at the number of memory channels per socket multiplied by the memory bandwidth per DIMM, we see a small jump as the memory frequency increases, but the larger jump is due to the increased number of channels. Going from DDR4 to DDR5 would have had a big impact on overall bandwidth as well. Looking strictly at memory bandwidth per core, a different picture unfolds, with Intel’s performance flattening out since 2019. AMD, on the other hand, shows a lot of change with its focus on core count. EPYC has more cores than Xeon, but memory bandwidth remains the same.
These findings may influence your decision on which vendor to use for your upgrade. Xeon is still a solid choice if you care more about memory bandwidth than how many cores your app can use. On the other hand, if you care about processing power and don’t have to worry too much about giving cores memory-intensive work, you should seriously consider EPYC.
