People like to say "synthetic benchmarks have no relevance to the real world" to sound smart, but that isn't ever entirely true. They are all relevant to soem

People like to say "synthetic benchmarks have no relevance to the real world" to sound smart, but that isn't ever entirely true. They are all relevant to soem degree. Somewhere in the world is a real task that looks like any given synthetic benchmark. The danger is that the real world use cases that show some relevance with the synthetic benchmark may be more or less rare. In this case, there are whole suites of benchmarks and real applications that have been compared between the new Epyc's and the Xeons and the only time Xeon has been on part with Epyc that I have seen so far is AVX-512 specific workloads. It is possible there may be some workloads where the worse ram latency of Epyc is a penalty that cannot be overcome by the larger L3 cache size. That used to be an issue on the previous gen Epyc, but the infinity fabric is faster now and the L3 caches are twice as big.