TornadoVM

TornadoVM logo

TornadoVM is a GPU programming framework for Java (JDK 21+, including JDK 21 and JDK 25) that JIT-compiles Java bytecode at runtime to NVIDIA CUDA C, OpenCL C, and Apple Metal (MSL). It runs the same Java source on NVIDIA, AMD, Intel and Apple Silicon GPUs, integrated GPUs (Intel HD Graphics, Apple M1-M5, ARM Mali), and multi-core CPUs. TornadoVM is a plug-in to OpenJDK and other JDK distributions (GraalVM, Red Hat Mandrel, Amazon Corretto, Microsoft OpenJDK, SAP, Azul Zulu); it does not replace your JVM, it complements it.

TornadoVM provides four backends, and developers choose which one/s to install and run:

  • OpenCL C: dispatched through an OpenCL runtime.

  • NVIDIA CUDA C: dispatched through the NVIDIA driver API; emits CUDA C and compiles it to PTX via NVRTC at runtime.

  • Apple Metal (MSL): dispatched on Apple Silicon GPUs.

On NVIDIA hardware, TornadoVM also calls directly into cuBLAS, cuFFT and cuDNN, and exposes Tensor Core mma.sync intrinsics from Java — see the hybrid API guide.

What is unique about TornadoVM?

Beyond JIT-compiling Java bytecode to each backend, TornadoVM’s runtime provides:

  • Live profiling: utilities to profile and debug code on multiple platforms (see Developer Tools).

  • Batch processing: automatic split and batch processing for big-data applications.

  • Multi-device & multi-backend support: offload and run multiple kernels concurrently across different accelerators and backends.

  • Multi-vendor: NVIDIA, Intel, AMD, Apple, ARM, and RISC-V hardware accelerators.