Microsoft Corporation has announced the creation of one of the five most powerful supercomputers in the world.
The company announced this at its conference for Microsoft Build developers, which was first held in a live stream format.
The new supercomputer was created exclusively for the non-profit research organization OpenAI. Its founders are the head of SpaceX Elon Musk and former head of the venture fund Y Combinato Sam Altman.
According to Microsoft calculations, based on publicly available data, this machine is one of five publicly announced supercomputers.
Its technical specifications are as follows: 285,000 processor cores (CPU core), 10,000 graphics processing units (GPUs), and network bandwidth of 400 gigabits per second for each GPU server.
The Microsoft supercomputer is designed based on the Azure cloud. As a result, it offers the benefits of modern cloud infrastructure, including rapid deployment, reliable data centers, and access to Azure services.
The machine will process data for different models and AI algorithms. The supercomputer will train the next generation of massive distributed AI models capable of performing a wider range of functions and requiring significantly more resources.
As Microsoft explains, historically, machine learning experts have developed small AI models that use marked data sets to learn how to perform a single separate task – recognize an object, translate text, etc.
But artificial intelligence experts have proven that some of these tasks are better performed by one large model, trained, for example, on billions of pages of publicly available text. Such a model has such a fine understanding of the nuances of speech, grammar, and context that it is able to perform several tasks perfectly: summarizing long speeches, moderating chat in online games, etc.
The creation and training of such massive models requires supercomputing infrastructure or clusters of modern equipment connected via high bandwidth networks. Microsoft believes that the new supercomputer for OpenAI is the first step to making the next generation of massive AI models and the infrastructure necessary to train them available to a wide range of organizations and developers.