News

Microsoft and OpenAI Unveil Supercomputer for Large-Scale AI Modeling

• By Kurt Mackie
• 05/28/2020

Microsoft took the wraps off a new supercomputer that it developed with, and exclusively for, OpenAI, an artificial intelligence company based in San Francisco. Hosted in Microsoft's Azure datacenters, it was created to train extremely large AI models. Microsoft made the announcement during its recent online Build conference.

Microsoft claims the new supercomputer's specs make it one of the top five publicly disclosed supercomputers in the world, and that it would qualify for the TOP500 supercomputers list. The as yet un-christened supercomputer is a single system with more than 285,000 CPU cores, 10,000 GPUs, and 400 gigabits per second of network connectivity for each GPU server, Microsoft said.

OpenAI was originally founded as a non-profit open-source organization by a group of investors that included Tesla founder Elon Musk. Today it comprises two entities: the non-profit OpenAI Inc. and the for-profit OpenAI LP. Microsoft, which is OpenAI's cloud services provider, invested $1 billion in the company last year.

The new supercomputer is being used to build large-scale models for things like language translation, text search, speech recognition, and computer vision, according to Kevin Scott, Microsoft's chief technical officer. The companies are finding that a single massive model tends to work better for certain tasks, such as learning by scanning text. Models only need to be trained once when they have massive amounts of data and supercomputing resources, according to Microsoft.

Turing Models Going Open Source
Microsoft has its own large-scale AI models, which the company is using to improve language understanding tasks across Bing, Office, Dynamics and other productivity products. These models are called Microsoft Turing models.

The Microsoft Turing models will be released soon as open source code, Microsoft said. The company said the Microsoft Turing AI model used for language is the world's largest, with 17 billion parameters.

In related announcements, Microsoft announced that it has updated DeepSpeed, which is an open-source deep learning training optimization library for PyTorch. PyTorch is an open-source library used for natural language processing and computer vision. 

Microsoft also added support for model training to the ONNX Runtime, enabling models to get trained 45 percent faster, the company said. Microsoft uses the DeepSpeed PyTorch library and the ONNX Runtime to train its AI models.

AI Bias-Checking Tools
The Azure Machine Learning service now includes "responsible machine learning" enhancements designed to help developers "understand, protect and control their models," Microsoft announced.

The InterpretML toolkit can provide such insights into AI models, Microsoft claimed. Scandinavian Airlines is using Microsoft's InterpretML toolkit to check the parameters used to detect fraud in its frequent-flyer EuroBonus program, for instance. 

Microsoft also has a new WhiteNoise toolkit for assessing privacy when using the Azure Machine Learning service, which is now available for developers as open source code at the GitHub repository. It enables analysis while protecting private information, Microsoft asserts. The WhiteNoise toolkit was developed in conjunction with researchers at the Harvard Institute for Quantitative Social Science and School of Engineering. 

In June, Microsoft plans to add a new Fairlearn toolkit for use with the Azure Machine Learning service. Fairlearn is another tool for developers that can be used to "assess and improve the fairness of AI systems," such as checking on models that recognize skin-tone, gender or age visual characteristics. The Fairlearn toolkit was developed by members of the research group Fairness, Accountability, Transparency and Ethics in AI (FATE). 

The advisory services company EY has used the Fairlearn toolkit on "a machine learning model the firm built for automated lending decisions." The toolkit is used to assess fairness based on gender-recognition capabilities, Microsoft indicated.