I have been running a witness node since September 2017. I have never put any shared_memory on RAM, all of it on a regular SATA SSD, and my RAM usage hovers under the 1 GB mark. Never had any problems at all, with disk usage around the 5% mark. Replays probably would take a while, but I have actually never replayed - not even once - since September 2017. It has just been rock solid. I've been playing around with a NVMe RAID0 server, and experimenting with /dev/shm, but can't think of a single reason to go the RAM route. Replay finished from scratch in three and a half hours, and as demonstrated above, replays are pretty rare. So, I'll stick with NVMe SSD for now, Optane next. Hopefully, with all of the scalability improvements, I won't be using RAM for years to come.

PS: I have noticed shared_memory is very compressible. I'd be looking forward to some compression tech built into Steemd. Of course, we can use existing workarounds now.