I'm thinking there could be contention on some lock somewhere. Can you attach the benchmark script you're using so I can try to reproduce it?
pgsql-hackers by date:
Соглашаюсь с условиями обработки персональных данных