the video demo shows inferencing in less than 5s - what kind of hardware was that done on? what are some sample inference times and benchmarks?