The Epoch of Reionization (EoR) 21-cm signal is expected to become increasingly non-Gaussian as reionization proceeds. We have used semi-numerical simulations to study how this affects the error predictions for the EoR 21-cm power spectrum. We expect
$SNR=sqrt{N_k}$ for a Gaussian random field where $N_k$ is the number of Fourier modes in each $k$ bin. We find that non-Gaussianity is important at high $SNR$ where it imposes an upper limit $[SNR]_l$. For a fixed volume $V$, it is not possible to achieve $SNR > [SNR]_l$ even if $N_k$ is increased. The value of $[SNR]_l$ falls as reionization proceeds, dropping from $sim 500$ at $bar{x}_{HI} = 0.8-0.9$ to $sim 10$ at $bar{x}_{HI} = 0.15 $ for a $[150.08, {rm Mpc}]^3$ simulation. We show that it is possible to interpret $[SNR]_l$ in terms of the trispectrum, and we expect $[SNR]_l propto sqrt{V}$ if the volume is increased. For $SNR ll [SNR]_l$ we find $SNR = sqrt{N_k}/A $ with $A sim 0.95 - 1.75$, roughly consistent with the Gaussian prediction. We present a fitting formula for the $SNR$ as a function of $N_k$, with two parameters $A$ and $[SNR]_l$ that have to be determined using simulations. Our results are relevant for predicting the sensitivity of different instruments to measure the EoR 21-cm power spectrum, which till date have been largely based on the Gaussian assumption.
