You could use a statistical approach by creating some random (or regular) point sampling set. You then assign the TIN or IDW DEM (Digital Elevation Model) raster values to this point set. You can then use a subtraction between both extracted values pairs and check standard deviation or variance between them. You could also have a try with a Pearson comparison (in Excel for example) to check if the 2 DEM's values are correlated (they should be).
You say "which kind of beats the point of the exercise." I don't agree : in my opinion, it doesn't seem problematic to rasterise the TIN, as long as you use the same step for rasterisation as the other DEM pixel size. They must be aligned. I would be more extreme, because I actually don't see how you can compare a vector set to a raster set without choosing at least a common data format. Either you vectorise you raster, either you rasterise your vector data.
In any case, you first assumption of substracting both is logical.