Expanding choices for PowerAI developers

We just published an update to our popular Power AI distribution! This updated release contains several new features and improvements to existing Deep Learning frameworks optimized for Power servers. Similar to previous releases of PowerAI tool kit, this release supports Power Linux platforms running Ubuntu 16.04 and CUDA 8 and was designed to take particular advantage of CPU/GPU NVLink on the S822LC for HPC, the world’s fastest Deep Learning enterprise server.

The major updates in this release are:

• TensorFlow, Google’s Deep Learning framework
• Bazel, Google’s build tool used for TensorFlow and many TensorFlow models
• DIGITS 5, the most recent version of Nvidia’s Deep Learning GUI. 
• NCCL, an optimized communications library for GPU computing
• Nvidia Caffe 0.15.13, an updated version of Nvidia’s Caffe fork

With this release, we’re adding the Google’s TensorFlow Deep Learning Framework to the PowerAI suite reflecting its growing popularity in the Deep Learning community.   We’ve also added Google’s Bazel build tool to this distribution, which is used to build many TensorFlow models, and of course, TensorFlow itself.  When installing TensorFlow, ensure that the underlying Ubuntu installation is up to date, as TensorFlow depends on a bug fix recently released by Canonical. (For those interested specific release levels and a detailed description of the defect, please refer to to Launchpad for a discussion of the defect and required release levels.)

This updated release includes the first version of NCCL, an enhanced communications library for GPU computing. The new Nvidia Caffe 0.15.13 has been built to use this new communications library. In addition, Torch may be configured to use NCCL by installing a Lua rock package with NCCL bindings. The PowerAI installation notes contain detailed instructions for enabling NCCL for Torch.

To simplify data management and training tasks for image classification, image segmentation and object detection, this updated release of PowerAI includes the DIGITS 5 GUI server. DIGITS 5 supports browser-based data preparation, and model training. Nvidia's DIGITS site has more information about DIGITS.

TIP: If you’re upgrading from a previous version of the PowerAI distribution, follow the installation notes for downloading and configuring the local repository, then replace the final installation command “sudo apt-get install power-mldl” with the upgrade command “sudo apt-get upgrade”.

To get the best performance for Deep Learning workloads on S822LC for HPC and replicate, the PowerAI installation instructions now contain a section with detailed performance tuning recommendations in the PowerAI installation instructions.

Get started with PowerAI to develop cognitive applications on Power today and share how you are unleashing the power of deep learning to transform the future of computing in the comments section.