It never really occurred to me to equate an underexposed image that is pushed (in Photoshop, for example) to make a 'good' exposure and a high-ISO image under the same conditions. If I understand this correctly an underexposed image that is pushed to be lighter in Photoshop will have similar noise attributes as a high-ISO image with the same resulting exposure range. The only difference is what is doing the gain increase--the camera or Photoshop. The limiting factor remains the number of photons that are read at the sensor, whether post-processing does the gain increase or the camera does it with an ISO, it really doesn't matter. Higher ISOs provide a more agressive gain table, in a way. Eh??
Do I have that right?
Yes or no, depending on the actual camera and ISO.
Emil's advice above is very good, as generous consideration. I find it generally usefult to explain something from different angle of view for a better understanding. I use now the Canon 5DMkII for the explanation.
Down below are the noise graphs at the different *real* ISOs, i.e. ignoring those, which are only numerical derivatives, as they are never useful for the raw data. The horizontal lines represent a given level of noise; the intersection with the graph of an ISO value shows, how great the dynamic range can be with that level of noise. The difference in EV between two ISO settings at a given level of noise shows, how much you gain (or lose) when selecting this or that ISO.
For example if you say you are accepting up to 40% noise (see the examples for how 40% looks), then see that it intersects the ISO 3200 graph at -7.5 EV and the ISO 1600 graph at -8.5 EV. The difference between these is one full stop (like with a fake ISO). This means, that if you choose ISO 3200 over 1600 (and use one stop lower exposure), then you lose exactly one stop of the dynamic range. In other words, if you reduce the exposure but leave the ISO at 1600 (underexpose ISO 1600 by 1 EV), then the shadows will look equally good - or equally bad - but you don't run the risk of clipping one stop more highlights than with 1600 (because the increased ISO cuts off the top EV of the highlights, independently of the effect on the shadows).
In cleartext: don't use ISO 3200 and above with the 5D2, except if the embedded JPEG (the preview) is important, or if you are recording raw + JPEG.
Next step: ISO 1600 vs 800. The difference between those is roughly 0.7 EV, i.e. that is, what you lose when shooting with ISO 1600 vs. ISO 800 (where ISO 800 would be shot with one stop higher exposure). Turn it around: with ISO 1600 you expose by one stop less, but lose only 0.7 EV. The economics is not great, but if you can't efford the higher exposure and the potential of highlight loss is not a concern, then choose ISO 1600 over the underexposed ISO 800.
Now, ISO 800 vs 400: the difference is 0.5 EV. That's not a bad tradeoff for one stop lower exposure.
ISO 400 vs 200: roughly 0.25 EV. This means, that if the DR of the scenery is not very large, you can reduce the exposure practically without penalty.
ISO 200 vs 100: the difference is tiny. Use ISO 100 only if you really want to max out the dynamic range.
ISO 100 vs 50: the graphs conform, what we know already: there is no difference between these; ISO 50 is simply an overexposed ISO 100, there is no reason to use it; in fact, there is good reason not to use it: avoiding overexposure.
Now, all the above was for the case that you are not free to choose the exposure. If you can efford that, then use the lowest ISO (but not the overexposing one).
Note, that the differences between the ISO steps are the same with all the relevant noise levels with this camera and many others, i.e. it does not matter which noise level you accept, the considerations remain the same.