Microsoft develops a super computer equipped with 285,000 CPU cores and 10,000 GPUs in cooperation with OpenAI



At the online conference 'Develop 2020' for developers, Microsoft announced that

Microsoft Azure will provide a new supercomputer service built in partnership with OpenAI , a non-profit organization that researches AI. It is said that the built supercomputer has the performance to rank in the top 5 in the world.

Microsoft announces new supercomputer, lays out vision for future AI work-The AI Blog
https://blogs.microsoft.com/ai/openai-azure-supercomputer/


Microsoft teamed up with OpenAI to build a massive AI supercomputer in Azure | TechCrunch
https://techcrunch.com/2020/05/19/microsoft-says-it-teamed-up-with-openai-to-build-a-massive-ai-supercomputer-in-azure/


The supercomputer built by Microsoft in collaboration with OpenAI is specially designed to train OpenAI's AI models. 'It's the first step to making the next generation of large AI models and the infrastructure needed to train them available as a platform for other organizations and developers to build,' Microsoft said. I will.

This supercomputer has over 285,000 CPU cores and 10,000 GPUs. Furthermore, each GPU server is configured with a high speed network of 400 Gbps and is connected to Microsoft Azure, but it is said to be utilized as a dedicated resource for OpenAI.

In addition, Microsoft is proud to boast top 5 performance based on the Top500 project that sets the ranking of supercomputers. Although Microsoft has not disclosed the detailed performance of supercomputers, the processing power must exceed 23,000 T FLOPS per second to be in the top 5 in the supercomputer ranking.

AI such as image recognition and language translation has been implemented focusing on a single task, but the trend of future research is focused on developing a large-scale AI model that can execute multiple tasks simultaneously. It is being done. For example, Microsoft Turing, an AI language model released by Microsoft in 2020, is the world's largest natural language generation AI that learns 17 billion parameters and is a real-time conversation while processing language, grammar, knowledge, concept and context. Manage, manage thousands of sentences, and generate code on GitHub.


by Bob Mical

Kevin Scott, CTO of Microsoft, said, 'Natural language processing can do hundreds of exciting things at once and computer vision can do hundreds of exciting things. When we see a combination of perceptual domains, we will have new applications that we can't even imagine now. ”

`` Every time we learn a lot about the limits of what we need and all of the components that make up a supercomputer, OpenAI CEO Sam Altman said, `` If we could design our dream system, What would it look like? ”And Microsoft was finally able to build it.”

in Hardware, Posted by log1i_yk