"Write once, runs anywhere." but not when it comes to GPUs - or other hardware site acceleration. The Java bytecode engine and GPUs are not a perfect pair. No wonder a lot of effort is put into solving this problem.
The new Foreign Function & Memory API in Oracles's Panama Project and the new VectorAPI, 3rd preview and 6th incubator version now, are promising and important changes to the JDK. Above all - it promises a better integration with external libraries than JNI.
But also 3rd party approaches like TornadoVM promise to accelerate Java on multi-core CPUs, GPUs, and FPGAs in an elegant way.
Let us discuss what is the problem in the JVM, for what we can use the Foreign Function & Memory API, and which alternatives are on the market.
Not only machine learning, but also applications like video editing and audio processing benefit from lightning-fast float operations.