Linear speedup

Well, yes. As I mentioned in the article, it's a rough measure. The actual details of what is faster is going to depend strongly not only on your hardware, but also your particular problem and how well it can be automatically parallelized.

The basic existing benchmarks are definitely exciting, though, for exactly the reasons you point out: for problems that can be parallelized, it sure looks like Bend does an excellent job of doing that without much additional effort on the programmer's part.