Reference PT is tracing pure paths from a camera and then bouncing them around. It’s basically what you would do if you would want to render an offline movie, where you don’t want any hacks or shortcuts and given enough time it will converge to a perfect result. Kind of, as in real production there’s denoising, ray guiding, caches, DOF added in post process and similar tricks.
Real-Time PT is a hybrid where some things are handled through rasterization and ray tracing, and some parts are path traced. Usually primary rays are replaced with rasterization, translucency is handled using standard techniques like probes or SSR etc. And then there’s a GI and reflection passes which trace some paths, or maybe there’s unified indirect lighting pass which traces paths handling both GI and reflections. Through depending on the implementation this may vary greatly. For example real-time PT in Alan Wake 2 is a mix of baked lighting and tracing short paths.
All of this is not really different from terms like real-time GI, where we have probes relit at runtime (no tracing at all) on one end, and lets say something like Lumen, which is tracing rays from pixels each frame on the other end. Both are both referred to as real-time GI, while they do completely different things under the hood. So as always devil is in the details. Details are complicated and hard to explain, so if anyone asks I tend to use “Reference PT” and “Real-Time PT”.
Now, why X is faster than Y is hard to tell without digging into details of what’s actually happening there. For example in Lumen reflections you can enable multiple bounces and they will trace very long and complex paths, which is quite slow. In general Lumen is designed for consoles and should provide a good tradeoff for that HW, while NV solutions are designed for high-end GPUs and should provide a good tradeoff there.