Performance wise, they ideally should be the same.
There's a chance that the human written techniques are a bit less optimal, so it might be a performance hit if you do poor human written methods of autoboxing. But, if you really want to get to it, there's an equal chance that a human might write some sort of non-general solution which beats default performance. Such a solution would not be as flexible, and it would probably trade off computational complexity for memory (like a big lookup array).
Personally, I'd recommend taking some time to really view the larger picture. Optimizing a single line or two of code is almost never a good investment. Reducing the amount of work necessary in the entire program is more likely to get you your performance boosts.
Note that in the general case, the JVM didn't change with the introduction of autoboxing, just the compiler did. So the compiler is adding the same instructions as you would write out manually in the most common cases. The performance is measure in the JVM at runtime, and if it's the same bytecodes either way, there's no reason to expect a performance difference.
This just smacks of premature optimization, but if you think you can find a difference in time: do so with careful testing, and then realize that it may be different on different point releases, operating systems, etc. It's just not a clear cut win in any case.