I did a benchmark. Naive Haskell implementation took 28 seconds, heavily optimized / hinted Haskell version took 11 seconds. Naive C implementation took 0.3 seconds.
None of the other code you wrote uses a linked list, so I don't know why you used them for Haskell. You used vector types for most of them (which Haskell has, by the way). Pretty disingenuous.
I admit that the prominent use of lists in newbie Haskell tutorials could have mislead you, but you seem to know the difference in data types since you used vectors everywhere else.
They aren't linked lists, but you still used a linked list internal to your "iter" function, where you used a vector (or array) there in other implementations.
The code is so incredibly close to your SML code (where you also spell out that you are using the Vector type). I am just blown away by how matter of fact you were in your statement.
Please either change your benchmark to use the same data type as all of the other examples, or stop using it as an example of Haskell's performance.
For your life-array.hs implementation, you replaced the Grid with an array, but not the "surroundings" with one, even though you used Vector for both in your SML implementation.
Built with -O2 option: https://gist.github.com/3189306
Sorry if I came off a bit assholish. People have been saying that you need to be a Haskell expert to get good performance forever, and in my experience this just isn't true. I will admit that there is clearly a problem with how things are introduced, as it is common that new Haskellers think they can use lists for every thing. I am sure part of the problem is that this is one of the few data structures available from the Prelude.
The act of finding the correct strictness hints seems a bit like black magic.