Gustafson’s work is interesting but very controversial. I’m not enough of a floating point maven to comment on the mathematics of it. It seems clear that FP apps would need to be re-analyzed and most likely rewritten to take advantage of the format; the economics of that suggest very low and slow adoption rates, even with a hardware implementation available.
However, the Mill is specification-driven. To add Gustafson-numbers (GNs) as a computation domain and their operation set is pretty trivial on the software side -a day or two to get the specification machinery to accept them and the tool chain to handle them and produce code. Then somebody has to figure out what a given operation on GNs actually does at the bit level and code that into the sim – maybe a week or two. Getting from there to gates is more, but not intolerably so. The biggest hassle would be getting them into clang/llvm as a recognized (and properly constant-folded and otherwise optimized) data type.
Would GNs sell more Mills? Almost certainly not. However, certain very specialized needs might want GNs badly enough to NRE a machine with them. And if such a party cared about costs, we surely can produce a GN-supporting Mill a lot cheaper than they could be added to any other machine. Interested parties please apply here 🙂