Just read through the SKA keynote and my head hurts with the scale of the computing requirements of that project. Energy, memory bandwidth and compute necessary to make the telescope work is enormous. The ask of using "free cycles" to convert between number formats might be optimistic...