It's difficult to argue about realistic "ranges" without being more specific about what your "range" means. One standard deviation? Over what domain of applications?
That said, speaking informally, your range doesn't look realistic at all, with a midpoint of 7.5% improvement. The midpoint is, IMO, more likely to be something like 15% or even 20%. And it will be the rare application indeed that runs 5% slower. I'd guess that 90% of all applications will land in the 10-30% improvement area.
Market Data 
Markets 
AI
