Some people may have been confused by IBM’s decision to support TensorFlow 0.12 with its PowerAI distribution. That wouldn’t be surprising since many continue to judge the company chiefly by its enterprise business roots. However, while the TensorFlow decision helps to clarify where IBM is today and where it aims to be in the future, it also details the practical steps the company is taking to get from here to there.
Why PowerAI matters to machine learning and deep learning
The aggressive promotion of artificial intelligence (AI) is rife with buzzwords but AI-related core technologies are fairly straightforward. In order to support AI-related functions and tasks, like speech and pattern recognition, computer vision and text analytics, underlying systems are trained with machine learning and deep learning frameworks.
While the two are complementary, they’re also quite different. Machine learning encompasses the algorithms and processes that enable systems to support AI-based skills. In contrast, deep learning focuses on the exercises required to train or tune “neural networks” to perform AI-related tasks. That process consists of AI systems ingesting datasets with millions or billions of elements, and performing training “reps” millions or tens of millions of times.
If you think this is similar to the “10,000 hours” required to become an expert that Malcolm Gladwell mentioned in his 2008 book, Outliers, you’d be right on the money. But there are far from subtle differences in how efficiently and how well various AI systems perform these tasks. With the right deep learning framework and foundational systems, AI projects can achieve their goals far more quickly and efficiently than they would using less adept servers.
That’s where IBM’s PowerAI distribution come into the picture. Based on the company’s Power Systems S822LC for HPC servers and POWER8 processors, along with partner NVIDIA’s NVLink interconnect technologies and Tesla Pascal P100 GPU accelerators PowerAI is designed from the ground-up to perform highly robust, highly optimized deep learning tasks.
Going with the TensorFlow
Why is IBM interested in TensonFlow? That’s a good question. TensonFlow isn’t the market’s only deep learning framework and, in fact, IBM also supports other competing frameworks, including CAFFE, Chainer, Theano, Torch and NVIDIA DIGITS. But TensorFlow has attained remarkably enthusiastic popularity and momentum since Google released it in November 2015. Despite that late date, GitHub data shows that TensorFlow became the most forked project for that entire year.
Why is TensorFlow so popular? Different people say different things, including noting its depth and ease of use, Google’s marketing muscle and the framework’s likely longevity (due to the depth of Google’s financial resources and commitment. In the case of IBM, be sure to add in Google being one of the five companies that founded the OpenPOWER Foundation in 2013. Today, the Foundation’s 300+ members are developing new data center solutions based on IBM’s open source POWER architecture.
It seems likely that at least a few OpenPOWER members have their eye on developing AI solutions but IBM is already heading in that direction. Along with its PowerAI distribution, the company plans to deliver both Technology Support Services for software support for the PowerAI stack and Global Business Services for deep learning design and development. In other words, IBM should be able to capture solid practical, strategic, tactical and economic benefits by supporting TensorFlow 0.12.
Every IT vendor finds a balance between building solutions customers need today and crafting those they will require in the years ahead ahead. IBM’s sizable investments in advanced analytics have already allowed it to establish a leadership position in related markets with its Watson cognitive platform.
But many of those same expenditures, especially the ones that expanded the capabilities of IBM’s POWER silicon and Power Systems solutions are resulting in sizable AI market opportunities. Not every customer will choose to utilize TensorFlow 0.12 but the framework’s popularity and acceleration suggest it will become an AI platform of choice for many of IBM’s enterprise clients.
Should TensorFlow achieve that level of success, IBM’s decision to proactively support it with its PowerAI distribution will be remembered as a prescient bet that led to handsome payoffs for IBM, its OpenPOWER partners and customers.