Benchmarking AI in Java: GPU 400x faster than CPU with Deep Netts using Project Panama
At Devoxx Belgium 2025, we presented the results of a new benchmark that demonstrates just how far Java has come for AI workloads – especially with the latest improvements in Java 25’s Foreign Function & Memory (FFM) API and the Deep Netts AI platform running natively on GPUs.
Benchmark Setup
We tested VGGNet, a popular convolutional neural network for image classification, implemented entirely in Java using Deep Netts.
The goal: measure how much performance we can gain by moving computation from CPU to GPU using Panama FFM API
Environment:
- Java 25 – improved FFM stability and throughput
- Deep Netts AI platform
- Task: Forward inference on pre-trained VGGNet model
Why Java 25 Matters
Java 25 brings major improvements to Project Panama, especially around:
- Lower overhead in native memory access
- More stable and predictable performance under load
- Streamlined GPU interop – no JNI glue, more clean FFM-based code
- Better compatibility with high-throughput multi-threaded pipelines
In short: FFM API is now production-ready for high-performance AI.
Benchmark Results
Why This Matters
This benchmark proves what many Java developers have been waiting for:
Java is not slow anymore – it’s AI-ready and GPU-powered.
With Deep Netts and Project Panama, Java can now:
- Train and run neural networks on GPU
- Deploy models in event-streaming systems (Kafka, Flink)
- Operate at native speed – while staying safe, portable, and JVM-based
JNI was neither easy nor productive. Panama is both easy and productive.
Watch the Devoxx Session
Catch the full presentation and live demo from Zoran Sevarac on YouTube:
👉 Devoxx Belgium 2025 – Java AI on GPU with Deep Netts and Project Panama
What’s Next
The new GPU acceleration feature will be available in the upcoming Deep Netts release in November 2025.
