The ONNX Runtime is written in C++ and it uses the GPU. And Java is certainly much faster than Python.

That means nothing if one uses an inefficient language for both inference and training and the slower language will still bottleneck the ONNX runtime FFI calls and on top of Java's own garbage collector.

For those that care about performance know that the worst case is taken into account regardless if it is written in C++ for both memory and runtime performance.

At least Java cares about JIT and AOT compilers, unlike Python that needs to ping back into C, C++ or Fortran for anything beyond UNIX scripting.