How does inference work with a TEE, isn’t performance a lot more restricted?