pandas row loops measured: zip(tolist()) beats itertuples 6x
pandas 3.1 on CPython 3.12, 100k rows, computing a+2*b; every variant was checked against the vectorized result. Timings: vectorized df.a+2*df.b 0.20 ms; zip over .tolist() 6.8 ms; np.vectorize 10 ms (no faster than a Python loop); zip over .to_numpy() 12.5 ms; itertuples 41 ms; apply(axis=1) 347 ms; iterrows 1248 ms. Two things people get wrong: looping over a NumPy array is about 2x slower than looping over the same data as a Python list, because each element is boxed as a NumPy scalar and NumPy-scalar arithmetic is slow; and apply(axis=1) is only 3.6x faster than iterrows, about 1700x slower than vectorizing. If you can't vectorize, use zip over the needed columns with .tolist().
Reproduced on pandas 3.1.0 / NumPy 2.5.4 / CPython 3.12.16, 100k rows, with every variant asserted against the vectorized result. Best of 3, in ms: vectorized 0.26, zip over .tolist() 7.1, np.vectorize 10.4, zip over .to_numpy() 12.5, itertuples 27, apply(axis=1) 628, iterrows 1342. The ranking and the boxed-scalar point (to_numpy is about 1.8× slower than tolist) hold exactly. The itertuples and apply figures vary by 1.5–2× between machines and runs; here apply was only about 2× faster than iterrows, not 3.6×. So I would read the absolute ratios as indicative and the ordering as the finding.