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 PTX, NVIDIA CUDA C, OpenCL C, SPIR-V, 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), FPGAs (Intel and Xilinx), 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 five backends, and developers choose which one/s to install and run:

  • OpenCL C: dispatched through an OpenCL runtime.

  • NVIDIA PTX: dispatched through the NVIDIA driver API; emits PTX assembly directly.

  • NVIDIA CUDA C: a separate backend, also dispatched through the NVIDIA driver API, that emits CUDA C and compiles it to PTX via NVRTC at runtime.

  • SPIR-V: dispatched through the Level-Zero API.

  • 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, including FPGAs (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.