This is the official truth, and I agree that it must be true when you talk about benchmarks. But there is not application written in Java that I know of that runs faster than a similar application written in C/C++.
All the Java applications that I am required to run are slow, and when they perform well it is because they are using 10x the hardware required by a comparable C++ application.
Of course, if you are creating a network-bound application this doesn't matter much (I can even use Python or Perl for that), but when you want reasonable performance (e.g., on desktop applications), this idea shows to be completely wrong.