Simon is basically correct, from an extensional point of view. We know pretty well what the semantics of modern functional languages are, and they really are relatively small variations of one another -- they each represent slightly different translations into a monadic metalanguage. Even a language like Scheme (a dynamically typed higher-order imperative language with first-class control) has a semantics which is pretty close to ML and Haskell's.
From a denotational point of view, you can start by giving a pretty simple domain equation for the semantics of Scheme -- call it $V$. People could and did solve equations like this in the late 70s/early 80s, so this isn't too bad. Similarly, there are relatively simple operational semantics for Scheme as well. (Note that when I say "Scheme", I mean untyped lambda calculus plus continuations plus state, as opposed to actual Scheme which has a few warts like all real languages do.)
But to get to a category suitable for interpreting modern typed functional languages, things get quite scary. Basically, you end up constructing an ultrametric-enriched category of partial equivalence relations over this domain. (As an example, see Birkedal, Stovring, and Thamsborg's "Realizability Semantics of Parametric Polymorphism, General References, and Recursive Types".) People who prefer operational semantics know this stuff as step-indexed logical relations. (For example, see Ahmed, Dreyer and Rossberg's "State-Dependent Representation Independence".) Either way, the techniques used are relatively new.
The reason for this mathematical complexity is that we need to be able to interpret parametric polymorphism and higher-order state at the same time. But once you've done this, you're basically home free, since this construction contains all the hard bits. Now, you can interpret ML and Haskell types via the usual monadic translations. ML's strict, effectful function space a -> b
translates to $\left<a\right> \to T\left<b\right>$, and Haskell's lazy function space translates to $\left<a\right> \to \left<b\right>$, with $T(A)$ the monadic type of side effects interpreting the IO monad of Haskell, and $\left<a\right>$ is the interpretation of the type ML or Haskell type a
, and $\to$ is the exponential in that category of PERs.
So as far as the equational theory goes, since these languages can both be described by translations into slightly different subsets of the same language, it is entirely fair to call them syntactic variations of one another.
The difference in feel between ML and Haskell actually arises from the intensional properties of the two languages -- that is, execution time and memory consumption. ML has a compositional performance model (i.e., the time/space cost of a program can be computed from the time/space costs of its subterms), as would a true call-by-name language. Actual Haskell is implemented with call-by-need, a kind of memoization, and as a result its performance is not compositional -- how long an expression bound to a variable takes to evaluate depends on whether it has been used before or not. This is not modelled in the semantics I alluded to above.
If you want to take the intensional properties more seriously, then ML and Haskell do start to show more serious differences. It is still probably possible to devise a common metalanguage for them, but the interpretation of types will differ in a much more systematic way, related to the proof-theoretic idea of focusing. One good place to learn about this is Noam Zeilberger's PhD thesis.