Table 3.

Energies assumed for the estimation of the power consumption.

Parameter	Energy
E m	5 pJ
E int32−add	0.1 pJ
E int32−mult	3.1 pJ
E float32−add	0.9 pJ
E float32−mult	3.7 pJ
${E_{compare}}^{\text{a}}$	0.9 pJ
${E_{float32-divide}}^{\text{b}}$	26.3 pJ
${E_{float32-1-cycle}}^{\text{c}}$	3.7 pJ

E_m is half the cache energy estimated by the energy cost table for 45 nm at 0.9 V from Horowitz (45) for an 8 kB 64-bit wide cache access, as only 32-bit loads are considered in this study. E_{float32−add}, E_{float32−mult}, E_int32−add, E_int32−mult values are also from Horowitz (45).

^aAs a compare operation can be performed as a subtraction which is basically an addition, the same energy is assumed for the compare operation. ^bA benchmark on the Ambiq Apollo 3 Blue with ARM Cortex M4F processor measured by executing three different measurements with a loop containing float multiply, divide and both instructions to cancel out loop and other overhead, showed a factor of 7.1 of float division energy consumption compared to float multiply energy consumption. ^cOther floating-point operations that take one clock cycle are considered to consume as much as a multiply operation.