"What Gives"... is that it's completely unimportant. Your 2nd device may have hit on a result that came from a different batch. Either way, the first time the device runs with defaults, and the benchmark at that could have been one on power-save and the other on full throttle, few possibles of many. After 10-20 completions, the averages will settle. Major -but-, there've been reports at Berkeley that the Time to Completions adjusts with huge lag in... 7.0.64 and same continuing into the 7.1 development versions so they are aware it could be wrong. [Seeing same here, for example a task after 5 hours at 90% and projecting still 2 hours for the remaining 10%].

Separate "Max additional work buffer" (MaxAB) at 4 days is rather useless if crunching only at WCG. Here you only get a limited number of units per work fetch [don't remember what the current standing CPU WU number is]. As soon as the cache is over the MinB level (Minimum work buffer), fetching stops, never topping the buffer out by itself. Don't like it... put the "What Gives" question over at Berkeley ;>)