> CPUs have gotten much faster, but moving bits around in memory has not.
Would it then not suffice to just load 32 bit floats, convert them to 64 bits on the processor, and then store them again as 32 bits?
Essentially, we are 'compressing' our data in memory, whilst doing the computational steps with full precision.
There is the 80 bit "long double" type that contains excess precision over the 64 bit double type. But if you want to write code that's both, portable, and reliable you have to turn off a lot of optimizations and use 32 or 64 bit types with standardised behavior.
Would it then not suffice to just load 32 bit floats, convert them to 64 bits on the processor, and then store them again as 32 bits? Essentially, we are 'compressing' our data in memory, whilst doing the computational steps with full precision.