I don't read that as them "claiming 100k TPS". They identified serially having to write utxo set as the bottleneck and they estimated a high-end server could do 100k TPS on that problem.
I did read it as 100k TPS.. actually 1M TPS is doable on high end servers (todays architecture) and 100k TPS easily on todays desktop computers.
"On a two year old 3.4 Ghz Intel i5 CPU this could be performed at over 180,000 operations per second. On newer hardware single threaded performance is 25% faster."
Seemed like a complete refactoring to make lean tx processing to achieve higher TPS effectively.