Well unums in particular (especially the new type 3 "posits" ones) are very interesting and practical for hardware implementation. John Gustafson, the creator o

Well unums in particular (especially the new type 3 "posits" ones) are very interesting and practical for hardware implementation. John Gustafson, the creator of unums, gave the first talk on type 3 unums at Stanford earlier this month, and there are already multiple hardware implementation efforts (one of which I am indirectly overseeing). While many people have recently hyped up low precision floating point for things like machine learning, Posits provide greater dynamic range AND greater accuracy for very few bits (as low as four, but with good ML starting 8 to 10 bits). On top of all of that, it costs less area and power in silicon than IEEE hardware. Check out the talk: https://www.youtube.com/watch?v=aP0Y1uAA-2Y https://www.youtube.com/watch?v=aP0Y1uAA-2Y