Writing GPU-Ready AI Models in Pure Java with Babylon
summary
Demonstrates three concrete paths: load an existing ONNX model via ONNX+Panama bindings and run it on local runtimes (CoreML on a Mac with unified memory) — example FER model yields logits then softmax and reports nodes placed on CoreML; generate Java representations of model graphs with Project Babylon and an OnnxLift generator to tweak operators (small refactor yields a measurable speedup); and write GPU kernels in Java with the heterogeneous accelerator toolkit (dispatch to OpenCL/CUDA backends). Example: Viola‑Jones detects 25 faces in 52 ms on demo hardware.
AI-written summary · about
chapters (9)
- 0:00GPU-ready AI in Java
- 3:26Speakers and Babylon
- 4:07Approaches to GPU AI
- 7:44Demo: pre-trained model
- 15:12AI models in Java code
- 29:23Heterogeneous Accelerator Toolkit
- 33:43Parallel GPU code with HAT
- 40:22Demo: face detection in HAT
- 46:42Summary and future
related talks
44mJava for AI
49mJVM in the Age of AI: Babylon, Valhalla, TornadoVM and friends
41mEscaping the Cloud: High-Performance AI in your Browser
42mAI Everywhere– accelerate your development from PC to cloud
29mML/AI models in JavaScript environment
50mUnder the HAT: Empowering GPU Acceleration for Java
33mMachine Learning in the Browser: Fast Iteration with ONNX & WebAssembly
50mProduction-Ready GenAI with Open Models for Java Teams