IBM is keen on being one of the best solutions when it comes to deep learning and cognitive applications. Their developer toolkit for POWER systems, PowerAI, already included support for a number of popular and powerful different frameworks though perhaps one of the most popular solutions had been missing until now. TensorFlow 0.12 has been added as an option for those looking to leverage IBM’s POWER8 architecture for their deep learning projects.

TensorFlow for POWER8

TensorFlow was the missing link for POWER8 DNN’s

Perhaps IBM is best known to the majority through their development of Watson, the multifaceted AI, or cognitive, framework that’s been¬†utilized to help many different industries. POWER8 and their specific HPC servers that feature NVIDIA’s Pascal P100 as an accompanying co-processor, are very well-suited to deep learning. In fact, with most frameworks it seems that IBM might provide the fastest platform for training new DNN’s. This is helped by the use of NVLink for communication between the many P100’s and between the CPU’s as well. That enables their platform, the S822LC server specifically, can perform those tasks with a bit of a performance boost over typical servers.

TensorFlow is a framework that was initially developed by Google for their own solutions. It’s since been made open source and has been widely adopted by many researchers and application developers that want to have cognitive aspects to their apps. It’s extensible and very powerful. And also hardware agnostic when developing. As in, it can transparently allocate any resources that it detects. That means you can develop it on anything and your network will run on any hardware that the framework supports, which includes phone hardware and even Google’s own creation.

IBM’s PowerAI toolset now includes the likes of Chainer, Caffe, Theano, Torch and more. The only missing framework seems to be Deeplearning4J, which is expected in the near future. Java applications tend to run very efficiently on POWER8 as well, compared to the current x86 counterparts.