This is probably one of the fastest possible FFI solutions, but if you're already enabling experimental features, Project Panama [0] is already in preview and will likely become the dominant FFI mechanism on the JVM.
I have spotted a Java ONE talk from NVidia, apparently they have revisited their collaboration with Oracle and have ported their CUDA bindings to use Panama now.
The whole purpose of Nalim is to be faster, Panama FFI currently perform similarly to JNI.