in the olden days of gp-cpu this was certainly true. is it still true? do you just mean "latency overhead for setting up a single PCIe transaction is much lar

in the olden days of gp-cpu this was certainly true. is it still true? do you just mean "latency overhead for setting up a single PCIe transaction is much larger than flinging a cache line across QPI/etc"?