On Skylake-SP platform most larger scenes already have more than 10% memory bound, this number will increase a lot more on EPYC2, since the CPU-to-memory ratio is a lot larger(6ch for 28c to 8ch for 64c). Or to put it simple, EPYC2 is far more thirst of memory performance than anything else. So in most rendering scenario I tested it's very important to populate as much memory channel as possible.
Besides, some EPYC motherboards may experience issues if you only have 4 channels populated, like having even worse memory bandwidth than it should be(tested with AIDA64).
Just remember, unlike 10 years ago, CPU performance has increased by more than 50 times, more than 100 times if you consider SIMD, but memory performance hasn't improved nearly as much. Sandy-bridge tests, even Skylake-sp tests aren't going to tell too much about the situation.