Why can't Stockfish solve this one?

This taken from a game I played recently (5|0), I was white, to play:

I was analyzing a line when I came to this position. The obvious move for white is Qf4+, trading the queens and winning easily after ...Qxf4 gxf4 Kxf4 h3. Yet the top Stockfish option is Qd3+, evaluated at +2.47. Furthermore, Qf4+ is not listed as even one of the top 9 moves. Yet after entering Qf4+, the evaluation corrects itself and jumps to +51.18 (though it takes several seconds to do so).

What is Stockfish's issue? (Or is my computer just dying?)

Thanks, Jasper_C

EDIT: after leaving the analysis board open, Stockfish found Qf4+ after about 8 minutes of "thinking"... an incredibly high time for an engine of its computing power. Is it faster for anyone else?

Happens on my computer too.

Actually, Stockfish found Qf4+ as the best move in less than a minute. However, it is reported as +4.74 whereas it is actually supposed to be > +60.

in order to rectify this problem, you must take the blood of a spring lamb and sacrifice it to Zeus at his altar at dusk

Will a turtle be suffice?

This is because humans can understand some concepts of in this case the winning pawn endgame. We just know by studying endgame books that ist is very basic the two pawns can always defend each other on h3 and f2 becuase Kxf2 runs into h4 and queening the h-pawn and thus can the King pick up the b -pawn and then walk back and win simple.  

However a computer does not know anything, it calculate. And that means it must calculate all the moves this concept takes plus all the counter moves of black if they are any relevant or not.  The computer is just programed to do this. This is why the computer is much stronger in positions in wich the possibilities are high with a lot of branches on every move. In complex positions the computer is much stronger then we and it takes a blunder out of our analyse in just seconds. However in endgames positions that we just know and understand and are long in moves they just suffer a horizont of calculation and in this positions the humans still prevail. By the way there are many more positions that the computer just does not understands.

Here stockfish 8 finds it instantly (at 8th ply).

I don't see how any engine would not see it, if it's not tuned to make mistakes.

Well engine isn't exactly below 1600

Gave it for ONE minute to some engines on an old Win7 computer that was once fast but has 4 years now, here is how fast they found and how much they valued it :

Andscacs took 17 seconds rating 40, soon 80+

AnMon didn'd found after a minute, I stopped.

Fire6.1 x64 my favourite, didn't found neither.

Hermann 2.8 x64 found in 6 seconds with a low value of +3 and didn't seem to realise anytime soon (within a minute) it was really better.

Houdini 15a x64 found in less than a second rating it +3 but still took time to rate it better (+10 @ 1mn)

Komodo 9 64 took 6 sec rating +5 growing to +250 @38s

Ruffian 1.0.5 didn't found.

Rybka 2.3.2a 64 found immediately but rated at +5 and was unable to understand the strength of the move neither.

SOS 5.1 failed completely

Spike 1.4 found immediately but rated +2 then +16 @ 1mn

Stockfish 9 x64popcnt took 23 seconds but rated better at +50, soon +130 and found the MATE (yesss !) just before 1mn! (He said Mate in 29 moves then 28, then 27 @ 2mn). After 10 minutes Mate was found in 23 moves.

I don't have Stockfish 8 to make a try.

Conclusion :

Ruben's Right. It's not obvious for a engine. They have to check all the way for every first move, not being able to think like us from the human perspective where "Queen exchange is good because pawns can travel then", it has to test travel from every first move perspective. You can see when reading their thoughts.
Engines did sometimes poor performances, unable to find the move was better, and/or underestimating it's real value greatly.

Best results after one minute are :

1) Stockfish 9 (only one to find the Mate)

2) Komodo 9 didn't took too much time and rated high fast.

3) Andscacs was OK with fair time and not too bad rating

4) Spike 1.4, Rybka 2.3.2a, Houdini 15a and Hermann 2.8 choose the right move too and pretty fast but failed completely to rate it at it's value.

Other failed in a reasonable time of 1 mn.

ASMFish, recent personal build: 1.Qf4+ suggested after 4 seconds as +42-something, #41 after 52 seconds (Core i7 4790K using 6 threads and 3G hash). Certainly all modern engines come to something like that sooner, or later.