In many others, especially those requiring vector or matrix arithmetic (medical imaging, physical simulations, modeling, graphics, scientific computing, video games), there's simply no competition.
Not sure of that. Some scientists from CERN would certainly disagree... Pure array-based arithmetic in Java is pretty damn fast, especially if you care to avoid dynamic allocations. It can be also very memory/cache friendly.
They do have quite a few projects in Java and they published some of their Java scientific libraries as open source. See Colt project. Also the software processing data from LHC is at least partially Java based.
10 millions element array of doubles/ints/bytes takes almost exactly same amount of bytes in Java as in C++. Java can blow your memory if you do stupid things like wrapping every point 2D into a separate object, but you know, good Java programmers know this. Other than that, its memory usage is quite reasonable, including some room for GC (typically 20-30% of total is enough).
In many others, especially those requiring vector or matrix arithmetic (medical imaging, physical simulations, modeling, graphics, scientific computing, video games), there's simply no competition.