It’s not too surprising to me. If TLB misses cost your program to run 6% slower, wouldn’t you expect a 30% reduction in that to be a 2% overall improvement? Why would it be more? If you’re surprised that TLB misses are not so costly, consider that cache locality and CPU prediction are largely going to hide the cost of misses (cache locality means the miss rate is low to begin with, prediction reduces the cost of a miss). TLBs are also super complex involving a few layers of caching (TLB look aside, L1 cache of your TLB table, L2 cache, L3 and then RAM).