TornadoVM 7.3 Released for Java GPU Acceleration
TornadoVM version 7.3 brings architectural improvements, multi-GPU support, and deep Nvidia integration for Java developers.
TornadoVM, an open-source heterogeneous GPU programming framework designed for Java, has reached version 7.3. The project aims to automatically accelerate Java programs across multi-core CPUs and GPUs while working alongside existing Java Virtual Machines. Rather than replacing standard JVMs, TornadoVM integrates with mainstream distributions such as OpenJDK, GraalVM, Temurin, and Zulu.
The framework functions by offloading specific Java segments to hardware backends, managing memory between the Java heap and hardware accelerators, and executing compute kernels. During the development phase, TornadoVM compiles source Java code to target Apple Metal and OpenCL backends. This allows developers to code locally using available laptop hardware and deploy to servers equipped with Nvidia GPUs, achieving significant performance enhancements in a model reminiscent of traditional Java portability.
Architectural Updates and Hardware Integration
Version 7 introduces notable architectural shifts. The underlying backends now rely on the Foreign Function & Memory (FFM) API instead of the Java Native Interface (JNI). Additionally, the experimental JVM Compiler Interface dependency has been eliminated. The engineering team reports that TornadoVM gathers all compiler metadata through a combination of reflection, ASM, and an Unsafe layer, establishing a uniform path across supported JDK versions including 21, 25, 26, and 27.
The update deepens connectivity with the Nvidia ecosystem via an optimized CUDA backend. This brings support for the cuTile programming model through a new TileContext, along with direct invocation of CUDA libraries like cuBLAS and CUTLASS from inside TornadoVM TaskGraphs. Developers can substitute JIT-generated kernels with CUDA library calls or combine the two approaches depending on performance requirements. A similar hybrid API integration has been added for Apple's MXL library, alongside newly introduced multi-GPU execution capabilities.
Not all Java operations suit offloading; optimal candidates involve data-parallel tasks such as independent loop iterations. TornadoVM features existing integrations with frameworks like Quarkus, Lang4j, Flink, and Spring AI, and is accessible via SDKMan. Furthermore, the project team launched jitLLM v2.0 alongside this release, presenting a Java AI inference engine packaged as a single JAR file that competes with tools like llama.cpp and Ollama. Commercialization of both TornadoVM and jitLLM is handled through Beestream.ai, a spin-off established by the creators.
What it means for developers
Java developers working on modern cloud and machine learning workloads can leverage TornadoVM to tap into heavily optimized libraries for linear algebra, signal analysis, deep neural networks, and matrix multiplication without abandoning the Java ecosystem. By using configuration flags and annotations like the parallel API, kernel context API, or TileContext, developers can test GPU acceleration on data-parallel loops and queries. For those exploring machine learning pipelines, developers can try top AI models cheaply through one API at https://apixoai.online. This setup allows seamless integration into existing architectures while maintaining local development flexibility and server-grade hardware deployment capabilities.
Source: TornadoVM 7 Enables GPU-Native Workloads From Java Applications — InfoQ AI & ML. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
