Is there any data or reference available for the baseline All-Reduce performance of TPU(v3 or later), similar to nccl-tests?

[ckjung1987]

Hello all,

It seems that the All-Reduce algorithm implementation on TPU is done as follows.

https://docs.jax.dev/en/latest/_autosummary/jax.lax.psum.html

Is there any performance reference or result available for different msg_sizes when the implementation is properly executed?

Thanks in advance!