GPU-wide memory is not quite as scarce on datacenter cards or systems with unified memory. One could also have local executors with local futures that are `!Send` and place in a faster address space.