300ms for inference is "fast" now? not even remotely usable in realtime