As an order of magnitude point, my experience has been that a bunch of CPUs trying to xadd has a throughput bottleneck on the scale of once per 50 to 100 nanose

As an order of magnitude point, my experience has been that a bunch of CPUs trying to xadd has a throughput bottleneck on the scale of once per 50 to 100 nanoseconds. But even if you allow an entire extra order of magnitude, at one per microsecond, that's still 10000 over the course of 10 milliseconds which is plenty for this usecase, at least for now.