In supplement to Wayne's fine answer, Robert Kosara has a recent post on his Eager Eye's blog about the very topic, Data Display vs. Data Visualization. In addition to as Wayne mentioned the goals of real-time visualization vs. more static displays might call for differences, he also mentions that gauges aren't very good for displaying multiple values. This is summed up nicely in his comment,
What you want to know is, how fast am I going right now? How much gas do I have left? What your speed was five minutes ago, or how much gas you had in your tank three hours ago, matters little.
So here is any obvious contrast between the goals of data visualization versus car-gauges, we pretty much always want to see multiple data values! And circular car-gauges are certainly a poor tool to do that. Sometimes we don't want to see multiple values though (a few circumstances are given in this question on the GIS site, What is the point of standard symbology?). And so we might expect other rules to which we apply the data visualization techniques in such circumstances. The GIS post I mention uses very flashy symbols/icons for point patterns that attempt to incapsulate the nature of the event (and sometimes visualization techniques like blinking dots to focus attention).
What I find interesting is that the work of Cleveland on comparing angles is still pertinent to car gauges though, and hence we still might expect a linear scale for a car gauge to work better than the circular display. So I suspect there might be more historical context as to why circular gauges were chosen (they are compact?), and it certainly may be this historical inertia as to why they are popular.
This much be a popular topic in the thralls lately, as the Visual.ly blog just came out with a post on the topic as well, Speedometer Design: Why It Works. In there they give credence to some of the things gung mentions in his post that I am somewhat critical about in the comments, in particular how we develop a gestalt for identifying locations around the circular display.
I think I'm partially coming around to this notion. A circular display provides more visual distinction between general areas than does a linear one. For a general example, it is easier to quickly tell the difference between a needle pointing to 3 o'clock and a needle pointing to 12 o'oclock than it is to tell the difference between 15 and 12 on a linear scale.
I'm still not totally convinced though, and I say rubbish to the notion that acceleration is easier to distinguish on a circular scale (or even if it is information we need the dashboard to inform us about anyway) that the visual.ly blog post mentions. Just my opinion though, I'm not sure any of us have been citing directly pertinent experimental results on human perception. Cleveland's is a start, but not likely to give an entirely satisfactory answer to these particular circumstances.
That being said the multiple data values are still the main crux of the argument, circular displays aren't good for multiple data values.