Re: CXL == slower /tmp
I've been pretty skeptical about CXL since I keep seeing people pumping it for all kinds of scenarios that make no sense to me, which is a red flag in the tech world. I've seen people claiming that servers would lose most of their DIMM slots and use CXL for memory expansion because they'd save pins on the CPU and board space in the server, which is about the dumbest idea I've ever heard since pins aren't at a premium and server density per rack is limited by power/cooling not by board size.
I do think the suggested use in cloud is probably the one place it may make sense, because you want the ability to move VMs around between physical servers but you don't want to spend money on equipping every one with the most RAM it can hold just in case it has to run a few memory hungry VMs. Having a pool of shared RAM out there that can be reallocated along with the VM's move avoids having to max out every physical server's RAM or limiting that VM to only running on the subset that have that maxed out RAM.
Now how common that is in the real world I have no idea. I've never dealt with hyperscale clouds but enterprise level VM farms seem to deal just fine with having physical servers of different configurations and restricting certain VMs to only running on a subset has never proven to be a problem. Even if you don't want to manage with a list of servers, something rule based "VM x can only run on servers with available RAM >= y" would seem to do the trick. I have to imagine Amazon et al have some pretty good automated management software that learns how much memory a given VM needs and can migrate it appropriately.
Time will tell whether CXL is the advancement some claim or a solution looking for a problem.