Notice that the real cost of idling on an atomic bus is that memory transactions generally take hundreds of cycles. The whole point of caching was to try to alleviate this latency, but implemented naively we will get little benefit, since only one transaction can be done at any time. For example, in the time spent waiting on a write to complete we could easily have made a BusUpg request to make everyone drop their cache lines, which doesn't need memory access.
Notice that the real cost of idling on an atomic bus is that memory transactions generally take hundreds of cycles. The whole point of caching was to try to alleviate this latency, but implemented naively we will get little benefit, since only one transaction can be done at any time. For example, in the time spent waiting on a write to complete we could easily have made a
BusUpgrequest to make everyone drop their cache lines, which doesn't need memory access.This comment was marked helpful 1 times.