Now, in the months since the Threadripper 2990WX debuted there have of course been a number of benchmarks that have been run. In a large volume of the benchmarks, there could be as much as a 50% performance degredation putting it in line or even slower than the 2950X, which has half the cores. In Linux however, this was not prevalent.
What does this mean? Well, a large amount of people believe this to be because of how the memory is split between the CCX's. For those unaware, the Threadripper 2990WX has 4 CCX's, although only 2 of these has direct access to memory, while the other 2 have to use a pass-through. It has long been thought that this is the primary cause of the 2990WX's performance degredation in multi-threaded workloads. However, this has since been disproved when tested against the Epyc 7551 which, like the 2990WX, is a 32 core 64 thread CPU although with slightly lower clock speeds. Unlike the 2990WX, the 7551 has 8 channel memory split between it's 4 CCX's. Does it still get hit by Windows? Yes it does.
Let's dig deeper. Level1Techs has performed independant research on the matter and was able to find that disabling Core0 in CPU affinity while running the program fixed many of the issues, freeing up performance headroom. It should be noted, that in all scenarios (even those with low scores) the CPU was locked at 100% utilization. He has since disclosed the belief that this is down to how the threads are continuously swapped between the cores on the CPU.
Moving on, Level1Techs has worked alongside Bitsum to help develop a tool called CorePrio, which effectively helps to eliminate the "NUMA Disassociation" (which is the term they have coined the bug with). This tool is available here for download, though it should be mentioned it is still being worked on.