The usual recommendation is to use epochs that are much longer than your planned analysis window, so that if any artifacts do occur they will be far away from the components you are interested in (this is not only relevant for ICA, but also for other operations like filtering, which can introduce artifacts near the edges of an epoch). For example, I often analyze a window like -200 to 1000 ms, but epoch the data from like -1000 to 2000 ms. (Of course, the epochs should not be too long; if they're so long that they overlap with the preceding and/or following epochs, you will have some data points represented twice in the dataset.)

Epoching is not always necessary for ICA, but to do ICA you usually should remove periods of data that are going to have totally different data than the ERPs you are interested in. For example, if you have people doing a cognitive task, but several times in the experiment they take a break and walk around and eat snacks or whatever, those break periods should be removed. Epoching the data is a quick and easy way to remove those off-task periods from the data. But if you're not comfortable with epoching before ICA, you could instead just manually remove the bad portions fromt he continuous data, and then run ICA on the continuous data.

The usual recommendation is to use epochs that are much longer than your planned analysis window, so that if any artifacts do occur they will be far away from the components you are interested in (this is not only relevant for ICA, but also for other operations like filtering, which can introduce artifacts near the edges of an epoch). For example, I often analyze a window like -200 to 1000 ms, but epoch the data from like -1000 to 2000 ms. (Of course, the epochs should not be too long; if they're so long that they overlap with the preceding and/or following epochs, you will have some data points represented twice in the dataset.)

Epoching is not always necessary for ICA, but to do ICA you usually should remove periods of data that are going to have totally different data than the ERPs you are interested in. For example, if you have people doing a cognitive task, but several times in the experiment they take a break and walk around and eat snacks or whatever, those break periods should be removed. Epoching the data is a quick and easy way to remove those off-task periods from the data. But if you're not comfortable with epoching before ICA, you could instead just manually remove the bad portions fromt he continuous data, and then run ICA on the continuous data.

You can run ICA before or after epoching. But ICA is known to work best with more data. Your continuous raw data gives more information for the ICA to work on to get you more stable components (ICs). Your epoched data has already lost some information due to pre-processing (windowing etc). So, ICA can only work on so much. Always remember “garbage in garbage out” principle.

I agree with both responses. Particularly as Akshay notes ICA is better with more data and if you can run it on the whole file you should do that. If you need to, or it is better for you to epoch it first then just do exactly as Stephen recommended and set larger windows.