In the perf lab, the benchmark is spending a considerable about of CPU cycles handling interleaving of the many worker threads, putting pressure on the OS scheduler: moving to virtual thread should help to improve our peak TPS making this cost to disappear (virtually. Let's say is greatly reduced).
Being a CPU bound kind of benchmark, saving CPU cycles is key for good performance but we need to make sure is a fair comparison w what Spring enable by default with JDK 25.