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