Thanks! His reports are exactly what I was referencing.
I tried to be realistic with the numbers I presented. The benchmark you linked is more proof that higher throughput per node is achievable, but none of what he did has made its way to the implementations (with reason I suspect, there are tradeoffs at each point)
A benchmark from Bottlepay again, a month or so earlier than this one, showed that an LND node running decent hardware (8vCPU, 32GB RAM) can do 33 payments per second.
I wanted to estimate this in a realistic way given that most of the network is running Pis, and ran with the conservative "4 tps" number, which again proves more than enough!