All I can do is say look at the math: change in x and change in y isn’t linear.
I know hologram already provided a good breakdown of the math on this but I'd just highlight that the method you're using to calculate change is flawed which is why you you are not seeing this as linear.
For one, as has been explained before, since we don't know what those power consumption numbers represent, there is no reason to assume the y intercept is 0. If the power consumption represents the SoC then setting GPU clock to 0 (essentially turning it off) doesn't mean that the CPU clocks are 0 which means that y(0) = 0 is not an accurate assumption. If anything, the conclusion you would draw from this would be that since y(0) =/= 0, yet R^2 is essentially 1, this would have to represent power consumption of the SoC or whole system when all other variables are held constant and thus it isn't just the GPU. But like, people here a lot smarter than me have already basically said don't read into these numbers so
Second. You shouldn't be measuring change in x versus change in y as a percentage for this comparison. As y values get bigger their relative percent increases are going to get smaller even if the raw difference stays they same. You need to calculate slope m using (y2-y1)/(x2-x1).
Doing so for 9.3W, 4.2W, 660Hz and 1125Hz you get m = 0.01096
Doing so for 12W, 9.3W, 1380Hz and 1125Hz you get m = 0.01058
Your slopes are basically the same. Any rounding in power consumption could explain the differences. There is also a margin of error at play.
As hologram said, if you did the linear regression R^2 being essentially 1 means the data is perfectly described by the model. Obviously 3 points does not make a statistically significant trend as also stated but there is no other data and it would be weird to have a perfectly linear consumption curve.
Last I would just say, all a linear relation says in practical terms is that a 1 unit increase in x (independent variable) should produce a proportional change in y along the line segment. It isn't about the percent changes because percentages are extremely sensitive to magnitude.
That curve is definitely linear.