The current dynamic advancements in AI, Computation and Connectivity, creates a platform for Smart Computing. However, how smart truly it is may be subject to debate from technological, business and costumer perspectives.
The ultimate metric for "smartness" is still the Turing test. ;-)
However, to answer your question as broad as you've asked for, I can only sketch a very general approach:
1. As you stated, it is a matter of perspective. As such, you need at first a metric of usefulness with respect to the perspective, or more precise, to the application domain.
E.g. a program to do, say,, matrix multiplication can be very useful in various very specific domains. Is usefulness may correspond to the speed or the maximum size of the matrices. However, there are a lot of domains where a matrix multiplication is not useful at all. In addition, the problem is so specific that most people wouldn't see (today!) such a program alone as "smart". That may change, if a software/system would provide a sufficient large number of specific functions.
2. A second component in being smart is a robustness or elasticity with respect to the ability to deal with aberrance's in the problem or lack of data.
E.g. in case of the software/system above, it would be usually regarded as much smarter if the human interacting with it does not need to state her/his problem is e.g. a matrix multiplication, but in terms of his/her experience model. Such robustness requires usually the ability to abstract (that is an ability that to generate the AI research was quite successful in the recent years)
3. Between both metrics (and probably within of each, since they are multidimensional themselves) exist a trade-off. The assignment of a performance depends quite a lot on how you want to solve the trade-off.
E.g., to what degree you are ready to accept that the system misunderstands your request, if the service in the cases it does understand your request performs well.
4. Further components can contribute to such a metric, e.g. the speed (response time) or resource need.
Also in our daily life, we frequently judge a person as less smart, if this person need longer time to perform a task, even it is often not a problem of smartness (but skilfulness, o.a.)
TL;DR: Such a metric would be multidimensional to a high degree. Only within very specific application domains such metrics could be useful and reasonable.