HAT Enables Java Offloading to GPUs

In January 2026, the Heterogeneous Accelerator Toolkit (HAT) and Project Babylon let Java developers offload Java code to GPUs and other accelerators using code reflection to generate GPU kernels. The article demonstrates matrix-multiplication optimizations achieving up to 14 TFLOP/s on an NVIDIA A10 versus 7 GFLOP/s on CPUs, and explains HAT's ND-Range, kernel, compute, and memory abstractions. This enables Java programmers to tune GPU workloads with near-native performance.