News for hackers. Stuff that matters.
Re:Memory Bandwidth Napkin Math by KiwiJohnno on Monday February 10, 2020 @07:37PM (#22294935)
If you are reading sequential addresses from RAM then the CPU cache is smart enough to prefetch what you want next, so you are actually hitting cache, not RAM. Remember, the "bandwidth" being referenced is limited not by actual bandwidth, but by latency. If you prefetch the right data, the latency dissapears.

Accessing a linked list (pointer chasing) means the memory accesses will be random, not sequential, so the prefetching doesn't really work.

There are no comments on this comment yet.