What's new in IBM PowerAI 1.6.0

Check here for the latest updates to IBM PowerAI information for the 1.6.0 release.

What's new

PowerAI is now packaged as Conda packages and no longer as RPM packages.

  • PowerAI Conda packages are available on the Internet in a Conda channel
  • There is no install package to download, instead connect to the PowerAI Conda channel and install your packages from there
  • Package dependencies are automatically resolved
  • Delivery of packages is open and continuous

CUDA, cuDNN, and NCCL are now packaged as Conda packages installed by PowerAI. Only the NVIDIA driver needs to be manually installed.

You can now run more than one framework at the same time in the same environment. For example, you can run TensorFlow and PyTorch at the same time.

Operating system support has changed:

  • Red Hat Enterprise Linux 7.6 is now supported. RHEL 7.5 is no longer supported.
  • Ubuntu 18.04 is fully supported

pai4sk now supports:

  • MPI based distribution of "data-loading, training, prediction and metrics" for snapML solvers
  • cuDF dataframes for snapML solvers in local mode
  • Accelerated ML algorithms from cuML

Updates for distributed deep learning (DDL) include the following:

  • Improved support for LSF
  • Support for TensorFlow Eager Mode

TensorFlow Estimator 1.13.0 is now included.

Previous tech previews are now included in base. These packages include:

  • ONNX
  • Caffe2
  • Large model support for TensorFlow, Caffe, and PyTorch

Technology previews

Packages:

  • TensorFlow Large Model Support (TFLMS) V1

Features:

  • Snap ML: snap-ml-spark library
  • pai4sk: Decision Tree and Random Forest algorithms from SnapML
  • RAPIDS.AI: cuML and cuDF
Note: For licence information for technology previews, see Technology Preview Code.

Package support updates

Deprecations

  • TensorFlow Large Model Support (TFLMS) V1 is being deprecated and replaced by TensorFlow Large Model Support (TFLMS) V2
  • Python 2.7 support: The Python core team plans to stop supporting Python 2 on January 1st 2020, ahead of this many deep learning frameworks and dependent libraries are or are planning to drop support for Python 2.7.

The following previously deprecated packages are no longer included with PowerAI

  • NCCL v1.3.5
  • libopenblas v0.3.3
  • powerai-protobuf v3.4.0
  • powerai-hdf5 v1.10.1
  • caffe-bvlc v1.0.0 (IBM-optimized Caffe is still included and not deprecated)