64 bit floats compiled with 32 bit compiler on 64 bit OS.

Question

I hope this has not been covered before, but if i compile a 32 bit program in c++ that uses 64 bit floating point numbers (double), and run it on a 64 bit OS, will it still take as many clock cycles to move the 64 bit float to the cpu and back to ram as it would on a 32bit OS because its compiled for 32 bit. Or would it take less clock cycles to because the OS moves in 64 bit at a time, even though the program is compiled in 32 bit compiler. The reason i ask is because im using VS express which has 32 bit only, and im wondering if i can use 64 bit floats while maintaining speed or if 32 bit floats will be faster, even though im using a 64 bit OS, and trust me, the program that i want to write will use tens of thousand of floating point numbers that will have many calculations and bit wise operations performed on them (looking into neural networks).

Thank you.

Answer 1

The 32 vs 64 bits you're hearing about is how many bits are in the address. It has little to do with how many bits are used to represent a double. In particular, 32-bit programs still represent a double in 64 bits, and modern processors have hardware that can process 64 bit floats natively (even if they can only process 32 bit integers natively).

So to answer your question, no it shouldn't matter. The speed of floating point operations should not depend on the 32 or 64 -bitness of either the OS or the compiler.

Answer 2

Doubles are slower than floats as a general rule.

A) it takes longer for the data to be moved in and out of memory. Busses are really wide nowadays, but it still takes 'more' time to move 'more' data.

B) Doubles take more time to compute, since the hardware tends to operate on groups of bits at a time when doing math, not all at once.

C) 4byte floats have had more use and therefore, more gates thrown at them. Note that SSE on x86 chips operate on 4 floats at once, usually in 4 clocks per instruction, There are some SSE instructions that use doubles, but only two at a time,

However, if you need the extra precision, then use doubles, and eat the performance cost.

You may want to look into GPGPU for crunching large datasets, Only the latest hardware does doubles, but their genearly GPU float performance is outstanding.