The idea is that, while a GPU may provide hardware-accelerated encoding/decoding - so the particular video job is running on an ASIC (not on a CPU), CPU is still orchestrating the whole process. So there's some CPU load and some latency.
For now, it should basically lower CPU usage in GPU-heavy tasks and make them more independent. So, for example, you're watching Netflix on an Intel CPU. The actual CPU usage will be small 5% (~3% on my 4 cores, 10 Mbps stream).
But actually there are a lot of tiny CPU jobs going on all the time, while the only crucial thing for the CPU should be feeding the GPU with video stream.
Also, it's probably a step towards functional multi-GPU systems.