[bri3d]

I don't think this would be good at all for most real workloads - you'd be taking the performance hit of having high-latency memory at all times. Even most hardcore NUMA vendors try to keep DRAM CPU-local, and writing high-performance software for NUMA generally involves ensuring that your data stays close to your CPU. Otherwise missing a branch or getting preempted by another task which flushes your cache lines becomes really, really expensive.

I do think this could be useful for a memcached workload, though, in tandem with some smaller amount of fast CPU local memory - you could basically share "memory bricks" between CPUs, and swap CPUs out independently without evicting an entire system's worth of memcache.