TornadoVM 2.0 is out today as the newest feature release for this OpenJDK and GraalVM plug-in that allows Java programs to run on heterogeneous hardware. TornadoVM targets continue to be OpenCL, NVIDIA PTX, and SPIR-V compatible devices for a range of accelerator support for use from conventional Java code.
The TornadoVM API allows for task-level, data-level, and pipeline-level parallelism for taking advantage of modern GPUs/accelerator hardware from Java code. At Phoronix we have covered the TornadoVM progress the past several years while now TornadoVM 2.0 is christened.
TornadoVM 2.0 brings some execution enhancements, GPU-native support for INT8 data types for NVIDIA PTX and OpenCL, support for compressed oops “coops” output, and zero-copy TornadoVM native array type instances with shallow memory segments. TornadoVM 2.0 also adds support for FP32 to FP16 conversion across all supported back-ends. There is also byte and half-float arrays in local memory now supported for all back-ends.
There are also a number of bug fixes, including OpenCL and NVIDIA PTX specific enhancements, including CUDA 13 changes. Downloads and more details on today’s TornadoVM 2.0 release via GitHub.
