some of nvidia's pascal gpus are able to do half precision computations at double the rate of single-precision computations. if lower precision is acceptable, this could also result in using half the energy, if it is true that moving bits consumes most of the energy.
Firephinx
Saving 5 operations on a calculation by fetching from memory is not a good idea on an energy constrained system because the power needed to fetch from memory is much higher than just doing the operations. Also, it takes longer to fetch from memory if you are fetching from DRAM.
some of nvidia's pascal gpus are able to do half precision computations at double the rate of single-precision computations. if lower precision is acceptable, this could also result in using half the energy, if it is true that moving bits consumes most of the energy.
Saving 5 operations on a calculation by fetching from memory is not a good idea on an energy constrained system because the power needed to fetch from memory is much higher than just doing the operations. Also, it takes longer to fetch from memory if you are fetching from DRAM.