CPU 操作花费的时钟周期

Operation	Clock Cycles
Simple register-register op(ADD, OR)	1
Memory write	1
Bypass delay: switch between integer and floating-point units	0 - 3
"Right" branch of "if"	1 - 2
Floating-point/vector addition	1 - 3
Multiplication(integer/float/vector)	1 - 7
Return error and check	1 - 7
L1 read	3 - 4
TLB miss	7 - 21
L2 read	10 - 12
"Wrong" branch of "if"(branch misprediction)	10 - 20
Floating-point division	10 - 40
128-bit vector division	10 - 70
Atomics/CAS	15 - 30
C function direct call	15 - 30
Integer division	15 - 40
C function indirect call	20 - 50
C++ virtual function call	30 - 60
L3 read	30 - 70
Main RAM read	100 - 150
NUMA: different-socket atomics/CAS(guesstimate)	100 - 300
NUMA: different-socket L3 read	100 - 300
Allocation+Deallocation pair(small objects)	200 - 500
NUMA: different-socket main RAM read	300 - 500
Kernel call	1000 - 1500
Thread context switch(direct costs) C++ Exception thrown+caught	2000 5000 - 10000
Thread context switch(total costs, including cache invalidation)	10000 - 1 million

Infographics: Operation Costs in CPU Clock Cycles - IT Hare on Soft.ware

Infographics: Operation Costs in CPU Clock Cycles

http://ithare.com/wp-content/uploads/part101_infographics_v08.png