In CXL 3.0 there's a whole switched fabric allowing multiple hosts to connect to & share multiple devices. This makes the whole thing much more interesting, more like the unsaid former-competitor Gen-Z who folded & joined CXL.
Right now CXL 3.0 is still spoken of like a M:N host-device fabric, and that could be excellent. But also, the potential here is dead obvious: if there's a switched fabric in the center allowing memory mapped interconnect, it definitely obviously clearly simply stands to reason that this is a potential host-to-host rdma system. You'd have to be an idiot to miss this. It's so simple, so screamingly obvious. And it's one built into the cores, no NIC needed, with far lower overhead.
If and only if no jarheaded shitmonger patent troll ruins this alliance, this is the obvious way to build multi-cpu systems, to replace top-of-rack switches. But so far CXL folks are utterly unwilling & unable to communicate these objectives. They still speak of CXL as a host-to-device system. Which is stupid.
This seems like it would be amazing even in consumers devices... And I hope AMD/Nvidia/Intel don't artificially segment it to the server space, as they have a bad habit of doing.
Right now CXL 3.0 is still spoken of like a M:N host-device fabric, and that could be excellent. But also, the potential here is dead obvious: if there's a switched fabric in the center allowing memory mapped interconnect, it definitely obviously clearly simply stands to reason that this is a potential host-to-host rdma system. You'd have to be an idiot to miss this. It's so simple, so screamingly obvious. And it's one built into the cores, no NIC needed, with far lower overhead.
If and only if no jarheaded shitmonger patent troll ruins this alliance, this is the obvious way to build multi-cpu systems, to replace top-of-rack switches. But so far CXL folks are utterly unwilling & unable to communicate these objectives. They still speak of CXL as a host-to-device system. Which is stupid.