Upload & Sell: Off
Here are some comments on different aspects of noise.
If the dynamic range of a scene being photographed at high ISO fits the "usable" DR of the camera at that then there is little or nothing in the way of downside to using high ISO. If you want to capture details at both ends of a high DR scene then high ISO is best avoided because no one exposure can ever do justice to it. i.e. there is no "correct exposure" whether exposing to the right or not, but if you find yourself in this situation often enough then it will pay to have an appropriate up-market camera that offers greater DR at the required ISO (greater DR at low ISO is no help) and/or plenty of extra pixels to facilitate downsampling.
Some cameras are better than others when it comes to handling poor lighting. Apart from greater DR, some will let you set the shutter speed you need for adequately freezing subject motion and the aperture you need for adequate DOF, and still give you autoexposure by controlling the ISO. I'm not sure about the latest Canon models but Nikon has generally done better in this regard. Using fully manual exposure control also works but is less often convenient and almost certainly increases the risk of not making an appropriate adjustment when a lower ISO would have worked better in a subsequent shot. I guess the ultimate solution is to maintain your own concentration on what you are doing and why you are doing it.
Check out the relative performance of camera sensors by viewing the charts at www.DxOMark.com . That is, use the charts and not the single-number scores. In effect the charts show DR at different ISOs, but it may be labelled differently. e.g. it could be the number of bits needed to count the number of separate colours or tonal levels that can be displayed. Bigger is better. A line above another line represents a camera sensor doing a better job that another camera sensor. Sometimes, however, the lines cross and a sensor that is better than another at low ISO may well be worse than the other one at high ISO. The single-value scores assigned to each sensor do not reflect this. Also, a camera sensor may have more DR than another at a particular ISO but can show fewer distinct tonal levels or colours within that DR. This too is info that is only gleaned from the charts.
At any ISO there is a limit to how much detail you can pull out of the dark areas of an image. Beyond that limit the noise intrudes greatly on the image details. You can only avoid this by exposing that area more, but of course you then risk burning out bright details at the other end of the histogram. I have found that in general it is better to correctly expose at high ISO than to underexpose at low ISO and try to recover in processing. Different cameras offer more leeway but in general this guide seems to apply - especially in situations that are the equivalent of needing very high ISOs to make a good exposure. Go too far with underexposure and you get thermal noise areas that are dominated by a single colour, on top of the usual distributed noise.
Of course, to state the obvious, we might be able to use a tripod to facilitate using a slower shutter speed or we might be able to supplement the available light with our own. Both generally involve carrying extra equipment with us and so have their downsides. In other cases they are simply not viable solutions to the problem due to excessive subject motion or distance to the subject(s).
Less obvious is that DR is actually subjective and varies person to person because the dark end of the DR is defined by an "unacceptably high" amount of noise amongst the dark subject details. What is acceptable to you may well be unacceptable to me. Most reviews that measure DR work on the definition of the bottom end being where noise level = signal level because it can be measured easily, but at the stage where noise = signal the signal is getting pretty hard to see in amongst the snow and would hardly be acceptable to anyone. It's more likely some of us want the signal to be two or three or four stops above the noise level even in the darkest areas. That's where the variability comes in.
Most noise increases with the temperature of the sensor. You might have the opportunity to accumulate multiple short exposures instead of taking one longer exposure. This could be beneficial in hot weather, particularly if you can rest the camera between shots. Not all scenes allow this to work satisfactorily. Also, it might help to not have the rear screen on during the exposure as it can get warm and is very close to the sensor. There's little point in reviewing shots on the LCD screen while resting the sensor if your aim is to keep the temperature down. At the very least you could reduce the screen brightness.
This definitely helps to reduce the visible impact of high-ISO noise but it affects small image details too. If you are trying to print larger then you will reach a point where a better camera was clearly the best solution because the print is just not working. Since it's now too late for that you'll just have to put up with the noise or make the print smaller or make the viewing distance longer. Also, I find that the nature of the noise changes as the ISO increases - it becomes more clumpy instead of being pretty much uniformly distributed single-pixel anomalies. Those clumps don't disappear so easily even with downsizing.
Personally I prefer high DR to downsampling because I mostly use Lr and downsampling undermines the benefits of working with raw image files and the psuedo edit commands that don't alter the original file. To downsample an image I would have to create a new physical file for Lr to manage and it would be either tif (big) or jpg. Neither has the benefits of a raw file. In particular, if I add a change to the raw file edit list then that change will not be reflected in any downsampled image file that was produced previously.
It seems that Topaz DeNoise can handle noise differently at different scene brightness levels, which is a good feature. Years ago I used Noise Ninja and it would divide the noise into three different physical size bands and allow me to treat each one differently. At the time NN was very highly regarded because it managed to retain nearly all of the desired image detail while maximising NR.
Nowadays I have all the time in the world and still run out of it, so I tend to stick with the Lr noise reduction. Later on I'll re-post my Lr NR technique to help you get the best of Lr with practically any image. I've yet to find a better Lr technique as most boil down to suck it and see trial and error tweaking that doesn't really have any method in their madness.
As with downsampling, using external NR software with Lr generally results in a new physical image file.
It may be worth keeping an excessively noisy image as a reminder to go back and get another later on, with more appropriate gear or at a better time of day. You could just throw it away because it doesn't exceed your minimum acceptable standard but then you would be far less likely to benefit from it. Use the EXIF data to check the exposure settings and assess whether or not you simply messed up by not choosing a lower ISO. Sometimes I've ended up with a high ISO shot of a static subject at fast shutter speed. That can happen because the subject is usually moving but had paused, or because of brain fade between shots. You can see from the image whether or not flash could have been used to supplement shadow detail, or whether a tripod could have been used. You can put all this together to make a shooting plan for the next time you go to that scene. Or you could throw it away and forget about it and get caught out again later on.
Above all, keep in mind that a noisy photo of a well focused subject is nearly always superior to a clean photo of a poorly focused subject. Don't be afraid of noise when there is no other option. It's either that, or else don't bother taking the shot.