Half precision

Are there any plans to let people get to the hardware half precision formats (BFloat16 and Float16) in Apple silicon? I have a research interest in mixed precision solvers and would be very excited to stop doing half precision work with only software emulation.

The issue to follow for that is Hardware Float16 on A64fx · Issue #40216 · JuliaLang/julia · GitHub. Apple silicon for now only has support for Float16 not BFloat16. If you are trying it out I would recommend building julia from source with the lines mentioned in the issue commented out.


Thanks for aiming me at that. I’ve never tried building Juila from source, but the discussion on that issue looks like there’s progress.