A lot of the IO usage can be avoided by a better configuration of virtual memory settings during replay and/or sync. By changing these settings, I found I was able to get close to a pure RAM-based approach for the state file using just the NVME storage for the state file and comment storage.
I wrote a script called reduce_writebacks.sh (included in haf_api_node) to set the virtual memory parameters. I never did enough testing to determine optimal settings, just messed around till I got something that worked good enough, but it improves replay time more than 10% and stabilizes results across runs: https://gitlab.syncad.com/hive/hive/-/issues/556#note_220737
It also saves wear-and-tear on your NVME drives...