Well exactly how are you defining latency? Is it random accesses you want to test? Or TLB thrashing? The R per RV prange benchmark accesses memory randomly,

Well exactly how are you defining latency? Is it random accesses you want to test? Or TLB thrashing? The R per RV prange benchmark accesses memory randomly, but in a TLB friendly pattern (TLB thrashing real world workloads aren't particularly common). The M1 gets 30ns, the 5960X is 55ns and still climbing at the max array size. The 5950X does better on "full" random, although that's unclear if that's a faster TLB, or the testing with a different page size. With Linux for TLB thrashing workloads you can always use HUGE pages, not sure if the M1 hardware supports that though.