The company, formerly known as Facebook, is building a supercomputer to serve the metaverse. When completed, the AI Research Supercluster (RSC) will be one of the most powerful computers on the planet and the fastest computing tool dedicated to artificial intelligence.
The RSC will be able to interpret the online communications of tens of thousands of people in real time, but even before that, they will be taught the most difficult work of our time, which is content moderation. Not only will the tool be able to solve real-world robotic problems, but it will also lay the foundation for nascent metaverses.
Metaverse is a 100 percent multi-sensory 3D experience, requiring AI for human use
Jerome Brenty, Vice President of Artificial Intelligence for Meta, told The Wall Street Journal.
Machine race
Notoriety here, stock prices plummeting there, the ex-Facebook is going full steam ahead in the competition in which big tech companies build artificial intelligence.
OpenAI’s language intelligence, GPT-3, founded by Elon Musk, has shown that significant performance can be achieved by increasing the size of the relationships within the algorithm, called parameters, and the dataset to which they are connected. GPT-3 in these characteristics was seventeen times greater than its predecessor, GPT-2. The Megatron AI that Microsoft revealed last year is three times the size of GPT-3. Chinese researchers and Google have independently developed algorithms that operate with thousands of billions of parameters. And Meta is preparing to take the next step: it teaches RSC to run an algorithm of 1,000 billion variables.
OpenAI is working on a similar project in collaboration with Microsoft. The machine they are making will be one of the five most powerful computers ever. Meta has not released details about its hardware, but what is known is that it will be on par with the world’s fifth most powerful supercomputer, the Perlmutter, when it comes to simple performance.
RSC currently includes a 6800 Nvidia A100 graphics processor, and it actually performs machine vision tasks twenty times faster than its competitors (hey, bad content edit). When the device is fully completed, it will have 16,000 GPUs and can hold about 36,000 years of HD video, storing exabytes. You will learn about 16 terabytes of incoming data per second at a speed of 5 exaflop.
One of the hardest parts of creating MIs, like language intelligence, is the need for an unforgiving amount of high-quality data that dwarfs the entirety of Wikipedia. The big advantage of Meta is that it does not have such a problem, because it crawls the largest amount of user data ever. If that is not enough, direct user activity can also be linked to the RSC. Concerns have already been raised about the latter, saying that it is anonymised, encrypted data and that the supercomputer is not directly connected to the Internet anyway.
At the height of artificial intelligence
The world’s most powerful supercomputer, the Japanese Fugaku, has no graphics processors and performs 442,000 trillion floating-point operations (short for flop). Based on the less precise metrics used in AI, Fugaku performs a single exalop (i.e. it performs a million trillion operations per second). RSC can be made the most powerful MI supercomputer in the world by being several times faster than Fugaku in this mode. Of course, there’s competition in this, too: the Frontier supercomputer, built at Oak Ridge National Laboratory and three times faster than Fugaku, could be a serious contender.
RSC would definitely be a great machine, especially given that language intelligence efficiency, for example, is constantly improving. DeepMind unveiled Gopher’s model of 280 billion variables last year, but what was even more interesting was the 7 billion variable RETRO that was introduced at the same time, which only needed access to an external database for routing and was able to suppress algorithms that It operates twenty-five times many parameters and is completely transparent. along the logic.
In general, not only do you create very powerful machines for a particular task, but the task itself is optimized.
(Cover Photo: Rvik Anadol installation of “The Hallucinogenic Space: Metaverse” on October 8, 2021 in Hong Kong. Photo: Miguel Candela/Anadolu Agency/Getty Images)
