Can somebody give me a convincing argument about 64 bit float?

Speculation, it hard to get proper docs On hyperthreaded cpu’s uses different ALU’s for 32 and 64 bits floats so you get better HT yield since there are still a lot of other 32fp calculations.