XLA Backend support for third party AI Accelerator

[VimalWill]

HI people,

I'm trying to figure a XLA's back-end support for our AI accelerator and I was going through the documentation on the XLA site for developing custom back-end and our requirements falls as a scenario - 3 [ no CPU like behavior and no LLVM back-end support].

I'm asking for any insights or dedicated design documentation that anyone could suggest. Your reply would be greatly appreciated.

Thanks
Vimal William