Previous attempts to access GPU compute performance from Java were challenging and led to squandered performance gains. Specifically, constraints imposed by JNI inevitably forced data copies when communicating with vendor-native libraries, and a lack of tooling made extracting vendor-compatible kernel code from bytecode a daunting task.

To solve data-parallel problems in Java, developers need to express multi-kernel algorithms in Java and have these kernels efficiently exchange data with each other and with the JVM.

In this talk, we will introduce HAT (Heterogeneous Accelerator Toolkit) and demonstrate how HAT leverages recent and proposed Java enhancements (Panama and Babylon) to provide Java developers access to untapped GPU potential.

Make sure to check the JavaOne 2025 playlist.