FreeBSD kicks Linux in the tail for network performance, especially if you require support for firewall. Berkeley Packet Filters (BPF) run as a pre-complied filter rules inside the kernel so much faster than Linux IPtables it isn't even a contest (which is why, as of kernel 4.10, BPF is being integrated into the Linux mainline, though the current code quality is a bit nascent).
Do note that the Netflix example is a "one way transmit" of 90gbps data outbound from a file store (media cache). It is an impressively optimized application, but it is also much simpler than would be required on a firewall. They are able to enqueue traffic to the NIC at PCIe speeds and let the NIC send them as fast as possible.
Much like my response in a different thread, for a FW application you have to get the bits in, look at them, and then send them back out. In a purpose-built firewall, once each "flow" is established and a rule is applied they will handle further packets on the same flow (simplest case, a matching 5-tuple) through a "fastpath" that acts much more like a switch than a router. This is fairly difficult to emulate on a general purpose CPU. In the pure CPU-based firewall, regardless of how fast your PF engine works, you end up inserting a large inter-packet latency. At 1Gbe and lower speeds this might be a small performance hit, at 10Gbe it could be a 50% performance loss, and at 100Gbe it is absolutely devastating. I imagine this is what may have led to
@BackupProphet's comment "But if you are using PFSense, I guess high performance isn't a top priority".