This question has been asked before.
The issue is that HTM does different things than other networks.
The comparison that comes to mind is comparing a F1 race car to a dump truck. They both have about the same horsepower engine but they use them in very different ways - one goes very fast and one carries very large loads.
DL is very good at mapping one high dimensional space to another. Exploring and mapping these spaces is a long and computationally expensive process.
HTM is very powerful in mapping states and transitions from one state to another. It can learn these transitions in a single exposure - something far beyond the current abilities of most popular DL models.
I would pose an alternate question - how do current DL models perform on HTM problem sets?
6 Likes