Is there any documentation about FP8 arithmetic functions?

Accelerated Computing CUDA CUDA Programming and Performance
1

When I read through the CUDA Math API Reference Manual,
FP8 does not documented arithmetic functions.
It documents data conversion and movement function only.
Is there any document about FP8 arithmetic function?

Reference

2

In PTX, e4m3 is only supported in data conversion and movement functions and tensor core instructions. Maybe there are no basis arithmetic instructions.

1 Like
3

I’m not aware of any typical arithmetic functions (e.g. add, subtract, multiply) provided by CUDA for any of the FP8 formats, currently.

The primary calculation method for FP8 computation would be tensorcore. There aren’t any other calculation paths (exposed, currently)

1 Like
4

E5M2 is basically the higher byte of the FP16 format. With it conversion is easy.

1 Like
5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.