Meta predicts that its new AI computer will be the fastest in the world by mid-2022

but the first phase of the new AI . As soon as the AI ​​Research SuperCluster (RSC) was built later this year, the company believes it will be the fastest in the world of AI computing, capable of “making about 5 exaflops of a well-integrated computer.”

The company says RSC will help researchers create better AI models that can learn from billions of models. Among other things, these models can create even more sophisticated tools as well as “seamless analysis of text, images and videos together,” according to Meta. Much of this work is done by his vision for the metaverse, whereas it says AI software and software will play a major role.

“We hope the RSC will help us develop new AI machines that can, for example, power real-time translation capabilities for large groups of people, everyone speaking a different language, to coordinate research or play AR games together,” said the program manager technical Kevin Lee and software engineer Shubho Sengupta wrote .

RSC currently has 760 Nvidia DGX A100 processors with 6,080 GPUs. Meta believes that this repetition already exists among the world’s most advanced AI computers. Based on the original benchmarks, the RSC claims that it can, in comparison with the company’s previous establishment, operate 20 computer screens more quickly and the NVIDIA Collective Communication Library more than nine times.

Meta says the RSC is also able to teach major types of natural language transitions three times faster. As a result, AI models that determine whether “action, word or image is harmful or positive” (e.g., eliminating hate speech) can be quickly taught. According to the company, the survey has helped protect people from the latest social services such as Facebook and Instagram, as well as what is happening.

Along with developing the physical design and operating systems of the RSC, Meta stated that it should ensure the safety and security of security systems to protect the actual training data it uses. It states that by using real-time data from its production systems, instead of the data available to the public, it is able to make better use of its research, for example, to identify what is harmful.

This year, Meta plans to increase the number of GPUs in the RSC to 16,000. It says this will increase AI performance by more than 2.5 times. The company, which began operations in early 2020, demanded that RSC train AI models on data sets up to an exabyte size (equivalent to the 36,000-year-old high-resolution video).

“We hope such a gradual change will enable us to not only create the right types of AI for our existing services, but also to support new users, especially in transformation,” wrote Lee and Sengupta.

Additional machines are being built in the US. Too late at the Argonne National Laboratory Department is expected , while El Capitan supercomputer, who will the world’s nuclear arsenals, and to top 2 exaflops by next year.