f32 (and even f16) is critical for fast computation for deep learning, especially on GPU and like for games. F32 is more than enough for the majority of data (images, sounds, text, financial data).
However from a library writer point of view, declaring a type/data structure/function signature as T, float or float32 is not worth the change of Nim default. That is just wrote once. Besides, I see the following arguments in favor of changing the default to float32: * Speed/Memory in microbenchmarks. This probably have influence on people looking at "as fast-as-C or as-efficient-as-C" languages and might help Nim popularity * Consumer GPU have fast f32 and slow/emulated f64. * Usability: Currently if you want to declare a seq[seq[float32]] you have to use f32 at each subsequence: let s = @[@[1'f32,2,3],@[4'f32,5,6]]. Arguments against are: * It's better to have accurate result slowly than wrong results fast. Especially for debugging * It's a breaking change * Compatibility with existing C and C++ libraries. My take would be to have an easy to understand default, either: * "We default to 32-bit int and floats everywhere" or * "We default to the platform pointer size" And document it in the manual.
