The main pages suggests you can take java code and it runs faster if there are gpus present. I like implicit concurrency so looked further.<p>The documentation on using it suggest there's a parallel annotation which tries to build something, while the main API is at <a href="https://tornadovm.readthedocs.io/en/latest/programming.html#kernel-api" rel="nofollow noreferrer">https://tornadovm.readthedocs.io/en/latest/programming.html#...</a>. That's the traditional clunky kernel interface from opencl et al.<p>So I think this is roughly exposing the usual GPU programming model, but in Java instead of C++, with some early steps towards openmp/openmp c style concurrency available as well.<p>Interesting project but not what the marketing page got my hopes up for.
I wondered about the performance delta to "native" CUDA kernels, and was disappointed there are only benchmark comparisons to single-threaded Java on their website.