We got around this by writing our own GC management: https://deeplearning4j.org/workspaces https://deeplearning4j.org/workspaces We write our own GPU algorithm

We got around this by writing our own GC management: https://deeplearning4j.org/workspaces https://deeplearning4j.org/workspaces We write our own GPU algorithms, Java native interface transpiler (eg: we generate JNI bindings) as well as our own memory management. We've found the JVM to be more than suitable. Granted - we wrote our own tooling and had reasons we can't move (those customers are a neat thing most people don't think about :D) I understand why you guys did go though. Congrats on pushing the limits of the runtime.