search

deeplearning4j / deeplearning4j

Suite of tools for deploying and training deep learning models using the JVM. Highlights include model import for keras, tensorflow, and onnx/pytorch, a modular and tiny c++ library for running math code and a java based math library on top of the core c++ library. Also includes samediff: a pytorch/tensorflow like library for running deep learn...
http://deeplearning4j.konduit.ai
Apache License 2.0
13.54k stars 3.83k forks source link

Google Released Coral for TPU #8313

Open eix128 opened 4 years ago

eix128 commented 4 years ago

Hi , google has released Coral SOM You can checkout details: https://coral.withgoogle.com/products/som/

It has TPU inside ( hardware neural net )

Any feature will be avalible to use this hardware with DL4J ? With also spark support with distributed training ?

AlexDBlack commented 4 years ago

It's not on our current roadmap, but in principle, I think we could add support, at least to the subset of ops that the TPU provides however ASICs have a significantly different programming model than say a CPU or GPU, so adding support isn't like adding say ARM CPU support, there's a lot of work involved not just a recompilation targetting the device or anything.

As for Spark on TPU: no, that would never happen, at least on this particular device. 1GB RAM and 8GB storage is nowhere near enough for Spark. Generally these boards (unlike the full TPU) are for inference only, not training.

Do you have a use case in mind for distributed use of such devices?

eix128 commented 4 years ago

Well yes , but you can make it by adding , just simple interface. There is already C++ lib for that. Link for C++: https://github.com/google-coral/edgetpu

and you can add it by using jni or javacpp https://github.com/bytedeco/javacpp