Associativity for Magma

In the absence of any further information then, yes, you need to check every triple. There is a theorem (due to G. Szasz) which asserts that on any set with at least four elements, there is a binary operation for which there is exactly one non-associative triple. (In fact, there are such operations on three-element sets also; $10$ of them, up to isomorphism.)

A reference for the Szasz theorem is:

@ARTICLE{Szasz1953,
AUTHOR = {G. Szasz},
TITLE = {{D}ie {U}nabh\"{a}ngigkeit der {A}ssoziativit\"{a}tsbedingungen},
JOURNAL = {Acta Sci. Math. Szeged},
VOLUME = {15},
YEAR = {1953},
PAGES = {20--28},
LANGUAGE = {German},
REVIEW = {\MR{56575 (15,95d) 09.1X}},
}

I should add that I've not actually seen this paper. (I've never found it online, and I don't read German anyway.) However, the proof is not difficult. Suppose you have a set $S$ with four distinct elements $a$, $u$, $v$ and $w$. Define the binary operation $\cdot$ on $S$ by putting $a\cdot a = u$, $a\cdot u = v$, and $x\cdot y = w$, for all pairs $(x,y)$ other than $(a,a)$ and $(a,u)$. Then it is easy to see that $(a\cdot a)\cdot a\neq a\cdot(a\cdot a)$. It is then tedious, but completely elementary to check (case by case, as it were) that every other triple does associate.

Generally, checking for associativity can be computationally very difficult. There are no easy visual criteria on the multiplication table to discern associativity.

In Rajagopalan and Schulman "Verification of Identities" (1997), an algorithm is given that probabilistically checks whether a given operation $\circ$ on a set $S$ is associative. If the operation is nonassociative, the test detects this fact with with any desired probability $\delta$ in $$O(\kappa n^2\log\delta^{-1})$$ time, where $n=|S|$ and$\kappa$ is the time to calculate $\circ$ for one pair of arguments. A variant of the algorithm will generate a specific triple $\langle a,b,c\rangle$ for which $(a\circ b)\circ c\ne a\circ (b\circ c)$, when one exists, in time $$O(\kappa n^2\log n\log \delta^{-1}).$$

This algorithm works for arbitrary operations. Unlike Light's test, it works even for "noncancellative" operations, where the equations $x\circ a = b$ and $a\circ x = b$ may not have solutions for all given $a,b$.

The paper also shows that if the operation $\circ$ is cancellative, one can compute a small ($O(\log n)$)set that generates it in time $O(n^2)$, and then apply Light's test to deterministically verify associativity in total time $O(\kappa n^2\log n)$.