Floating point is already incredibly imprecise. The thought of debugging any piece of software on this system sends shivers down my spine.
I would much rather see something like UNUMS (https://news.ycombinator.com/item?id=9943589), which provide improved accuracy AND energy efficiency over IEEE floating point.