But why? What's the algorithm/architecture here? I am genuinely curios here.

I guess we can generalize finger/face readers as multichannel cameras. What do you gain by computing a "hash" of the data (and the associated machinery to send that hash securely) versus simply sending raw data for evaluation at CPU level? In the end the CPU has to trust the data sent by peripheral anyway.

I understand the use of separate compute unit when extraction of secrets must be protected, but in this case it is the CPU that protects those secrets anyway.