This could be very fast at running embedding models (if it supports the right modes)

Considering that memory bandwidth is the most important for inference for most neural networks, that matters more.