We are on the Windows platform with CUDA version 11.1 and the newest Nvidia GPU card. We have tried but found some trivial to overcome. Until now we have obtained nd4jcuda.dll from building libND4J from source. Since the most trouble parts are the layer between operating system with CUDA driver and the applications. We think the nd4jcuda.dll will do that for us. We wonder if it is neccessary to build other parts as DL4J, ND4J, SameDiff, DataVec, Arbiter from source or we can just download the compiled and released ones from any repository without any troubles so can avoid the trivials follow on. So what are the neccessary modules other than libnd4j must we build from source also?
@TempKonduitUser1 If you have libnd4j compiled, then that’s more or less all you need.
The only other cuda related components is deeplearning4j-cuda which has some of our cudnn components. Most of the library besides the native parts that involve javacpp (the c++ to java interop library we maintain) - so you only need to build what you’ll use. Most things should be just a straightforward install though. Could you be more specific about your intended environment?
@agibsonccc Do you mean the native parts that involve javacpp (the c++ to java interop library) are within libnd4j?
@TempKonduitUser1 yes, it’s one c++ library. The rest are libraries already pre compiled by javacpp.
Basically it goes:
java library → javacpp → libnd4j
We maintain javacpp as well. Libnd4j was originally built to mainly work with that before evolving in its own library. That should be all you need.