Update: I ran a quick benchmark (large matrix multiplication) between GNU APL, J, and NumPy (standard pre-compiled linux amd64 packages, all running single-core.) Here are the results.

GNU APL (1.7)

    ⍴+.×⌿?2 3000 3000⍴1e10

    - size 3000: 65s, 9GiB RSS
      (crash on bigger sizes)

J (8.07)

    $(+/ .*)/?2 3000 3000$1e10

    - size 3000: 2s, 0.2GiB RSS
    - size 10000: 60s, 3.7GiB RSS
      (crash on bigger sizes)

NumPy (1.13.3, blas/lapack 3.7.1)

    import numpy
    a=numpy.random.randint(0, 1e10, (2,3000,3000))
    print((a[0,] @ a[1,]).shape)

    - size 3000: 25s, 0.2GiB RSS
    - size 10000: (>15m, I killed it) 2.3GiB RSS

Conclusions:

J's implementation is surprisingly performant! Easily beating NumPy on speed alone by a factor of 10 or more! (And I thought blas/lapack were already heavily optimized libraries!) J's memory usage is comparable to that of NumPy. GNU APL had the worst memory and cpu profile of them all.