Feature Request: 8 bit fixed point numbers

[jenkspt]

I'm interested in using 8 bit fixed point numbers in jax and numpy in the range [0, 1], and [-1, 1]. I think with the current naming convention it would be something like float8_e0m8 and float8_e0m7.

[hawkinsp]

Well, I'm sure we'd accept a PR to add such a thing. Can you say more about the number format you want? Is there a paper or similar describing it?

In general we're adding types to this package only as we need them for some concrete purpose (e.g., there's a hardware vendor that offers them). What is the use case here?

[fayalalebrun]

@hawkinsp Fixed point formats should be usable with any hardware which supports integer arithmetic (source). Therefore the hardware support should already be there as far as I understand.

As for use case, integer arithmetic is far more efficient to implement in hardware as opposed to floating point . Therefore if researchers have such formats available to them, it could drive forward the development of more efficient neural processors. Moreover, something like this could be useful to FPGA engineers in the prototyping stage, where fixed point numbers are commonly used due to the aforementioned hardware efficiency concerns.

I could help with this, but would appreciate any pointers to the relevant numpy documentation. The types would just be wrappers over the regular integer types, with bit shifting applied accordingly. Though sin and square root might be an issue.