I'm exploring an idea to reduce GPU dispatch overhead in a Java-based runtime, TornadoVM, which executes compute operations from TornadoVM bytecodes.