After the serial-to-parallel to break input bitstream into groupings we do the constellation mapping into values of phase/amplitude (or I, Q). For 64 sub-carriers this gives us 64 pairs of amplitude(n)/phase(n).
an = amplitude(n) * cos(phase(n))
bn = amplitude(n) * sin(phase(n))
sub-carrier(n) = an * cos (n * delf * t) + bn * sin(n * delf * t)
Why do we need the IFFT. We could just compute the 64 sin/cosine time functions, sum them and then up-convert to passband. We lose the complex notation (an + j bn) where we keep the an and bn terms separate -- and do the up-conversion separately for both the an and bn sums.
At the receiver, we could just do a real sample to get 2*64 = 128 samples of the summed sub-carriers. These real samples could be input to an FFT to get back the an and bn terms. This is in contrast to separating out the an and bn terms at the receiver This would simplify the hardware.