I inferred that they meant the neural engine cores by neural accelerators or it could be a bigger/different AMX (which really should become a standard btw)