Couple thoughts...

It's spelled EPYC...

There were some weird NUMA considerations when trying to saturate 100Gbps Ethernet NIC's. I saw some firmware improvement/workarounds (basically stick TCP to a core so it wouldn't span NUMA nodes). 99% of people are not going to care about this, and the work around seems to be holding from what I've seen.

For Single socket it's really the best bang for buck out there. I've seen some impressive vSAN numbers on it if anyone's wondering. It doesn't lock out 1/2 the PCI-E or Memory lanes like Intel in a single socket config.

I've got a RYZEN at home, and quite happy with it.

I'm hearing Intel made some optimizations for multi-threaded and cloud native workloads in Skylake that cause issues for .NET and single threaded apps. I haven't done any bake offs but I'm curious if AMD went down the same pass (especially with spin locks). If you have older .NET code with lots of spin locks you want to get it updated or get ready for potential (heavy) regressions on performance vs. Broadwell. We are entering a weird situation where CPU vendors are starting to NOT optimize for enterprise applications but for the mega clouds.

While I think ARM is awesome, there also is a LOOOOOOOT of optimization to be done with compilers before it gets taken seriously outside of edge cases, or homogenous workloads where people will invest the time like CloudFlare.

It requires a outage to migrate between Intel and AMD at a virtualization layer. Intel also has QuickAssit and some other interesting offload, as well as Apache Pass technology.