Google TensorFlow Learns to Compete

by Diana Dehterevich, with Dan Mandell | 06/03/2019


TensorFlow is among the most popular open source frameworks for machine learning (ML) but its leadership is under siege. Produced by Google, TensorFlow has been adopted by numerous companies including Airbnb, Arm, Intel, Lenovo, Qualcomm, and Twitter. The platform is known for its easy model building, flexibility to provide models individually or together, and live training for customers, among other features. While many companies have released their own ML frameworks, not many have been able to keep pace with TensorFlow, which is provided for free by Google. However, competition is ramping in the space due to rising adoption of AI across industries, rapid growth of the AI hardware market, and increasing ease-of-use for ML tools and developing algorithms, and may lead Google to rethink its approach.

Google makes no profit off of customers using the platform, but it does keep the data required to train and build algorithms. The company values computing data over modeling algorithms, which can be done by many programmers. By focusing on having access to large datasets, Google collects the information, which the R&D team utilizes in their future production ideas. This has provided Google a competitive advantage allowing it to get ahead of other ML vendors in one of the most rapidly growing technology markets.

However, the company’s leadership in the market is under attack as other ML vendors are concentrated on improving their ML frameworks by making them faster, more efficient, and compatible with other platforms. For instance, PyTorch, developed by Facebook, stands out for its computational graphs. Meanwhile, Cognitive Toolkit (CNTK) by Microsoft emphasizes its speed and interface. Other rising competitors include Caffe, Caffe2, Torch, Keras, and MXNet. These frameworks are advantageous for their speed, image processing, and availability of numerous pre-trained models. Due to such a high competition among the ML framework developers, the release of Open Neural Network Exchange (ONNX) can be a solution for developers who cannot choose a platform. ONNX is an open format to represent deep learning models that allows developers to easily move models between different platforms. Even though Google decided not to participate in the project, TensorFlow is one of the supported frameworks.

Whereas the ML software competition is rising, the same can be said for AI chip production. Google has incorporated the data learned from offering TensorFlow to releasing the Edge Tensor Processing Unit (TPU) – an Application Specific Integrated Chip (ASIC) designed to run inference at the edge level. Other ML vendors are also realizing the necessity of developing ML hardware that can deploy their ML frameworks. For example, leading embedded semiconductor provider, Intel, acquired Nervana Systems, for over $400 million, which enabled it to release a Neural Network Processor (NNP). The company has recently entered into a partnership with Facebook on an AI-inference chip.

Even though Google put itself on the fast track to leadership in offering ML frameworks, the vendor has been gradually losing its top spot in the space for ML software due to growing competition from different innovative solutions. To retain its leadership at the top of ML frameworks, Google will need to provide competitive and scalable AI hardware as well as broaden support for third-party integrations and more-open development workflows.

View the 2019 IoT & Embedded Technology Research Outline to learn more about our coverage.



Back to Top