Sunspider is an ancient, awful benchmark suite that has been optimized to the point of insanity. For example, every major browser has a daylight savings offset cache that is no use to any real code but speeds up Sunspider significantly. It needs to die. You should ignore it.
You can delete the word "browser" from that sentence and it's still true. Just one example: You wouldn't believe how many large projects rely on Dhrystone for benchmarking and selecting silicon.
I can believe it. But at least in the C/C++/Fortran world there are decent suites, e.g. SPEC, even if not everybody uses them. The browser world doesn't even have that.
I think the brouhaha over Mandreel vs Emscripten glosses over a deeper issue, which is, the vast majority of Web apps aren't C++ compiled to asmjs, and performance on asmjs is not going to be a good proxy for general web performance.
"Real apps" using idiomatic JS should be used as benchmarks, and those apps should approximate apps in widespread deployment. For example, how fast can a reactjs re-render for a complex page run?
Octane is better, though still flawed: https://blog.mozilla.org/nnethercote/2012/08/24/octane-minus....
The state of browser benchmarking isn't good. Here are my ideas on how to fix it: https://blog.mozilla.org/nnethercote/2014/06/16/a-browser-be.... The political/organizational challenges are as big or bigger than the technical challenges.
And really, you should care most about how each browser performs on the workloads you are interested in.