For the record, this is the IPv4 NAT performance of a Mellanox SN2010. The fact that it only has a single interrupt/queue is the major limiting factor, it's maxing out one of the 4 CPU cores while the other 3 twiddle their thumbs.
(The upload performance is probably also limited by Firefox not being able to send enough.)
IPv6, being hardware line rate switched, of course has none of these problems…

