My gut feeling is that implementations of hardware acceleration are just simply not as good as software implementations, the cause perhaps relating to localised resource constraints in hardware when recursing deeper or somesuch.
The algorithms are the same, the chip versions are more resource constrained as the stack goes deeper if you will.
Hardware compressors are much much faster, no doubt, by an order of magnitude and more .. but software compressors left to run slow, with two passes, just produce a much better result.
Standard hardware compressor settings tend to leave moire patterns when the camera pans outside a chain link fence onto a player inside a court (for example), or leaves a blocky halo around the significant figure point of interest moving in an otherwise stillish scene - it's subltle small things that grate when you see them.
The root cause of perceived downgrade in quality is bulk pushing of high quality originals through hardware accelerated compression in order to save time and reduce cost when generating lower bitrate versions to stream.