Regarding the variable length of his FP numbers, their energy savings are beer-based numbers, I don't think they have any experimental basis (yet).
Also, because they are variable sized, you need to access them through pointers if you want random access. Now you are reading both the pointer and the value from memory. Since pointers and double precision floats have the same size on modern hardware, one would expect this to actually consume more energy than just reading a double precision value. An additional indirection can also have great performance cost. And there's one more step you need to do. After getting the pointer you need to first read the utag so you can decide how many bytes to read. So where you would normally just read the value, you now need to read the pointer, use that to read the utag and use the utag to read the value.
