Meta is building a supercomputer with the goal of implementing learning algorithms that provide him with the knowledge needed to continue creating his virtual world.
According to the company, The AI Research Supercluster (RSC) will indeed be among the most powerful devices on the planetAnd that’s not over yet, because once it’s built it will be the world’s fastest artificial intelligence computer.
The applications of the aforementioned supercomputer will be very diverse, initially, Meta expects that RSC will improve its products and social networks by training better algorithms that detect malicious content, and in the same way, the company will aim to improve communication around the world, because according to Zuckerberg, progress can What was made with the team is to enable real-time language translation among thousands of people online.
On the other side, RSC has its eyes on metaverses, pues según explicó Jerome Pessenti, vicepresidente de IA de Meta al medio internacional Wall Street Journal: “en el metaverso, el cien por ciento del tiempo es una experiencia multisensorial en 3D, y necesitas crear relevant agents de inteligencia artificial for you”.
Many companies are turning to more powerful AI models powered by supercomputers, because according to SingularityHub, OpenAI’s GPT-3 algorithm was capable of routing more than 175 billion parameters simultaneously, which is 17 times larger than its predecessor from the same company. Building on the success of OpenAI, Microsoft introduced the Megatron AI algorithm in 2021, which is three times larger than GPT-3.
Google decided not to be left behind and developed an algorithm capable of analyzing more than a trillion parameters and now, Meta has arrived announcing that it plans to use RSC to train algorithms with billions of parameters.
Supercomputers receive this designation because of their power, but they can have completely different uses and formulations., it all depends on the purpose for which it is designed, because it can be used to simulate physical phenomena such as the elementary particles of the Earth’s climate and the motions of the universe in general as well, in the field of artificial intelligence, as in the case of Meta and some of the most powerful machines in the world.
in this time, RSC runs on over 6,800 NVIDIA A100 graphics processing units (GPUs)What’s curious is that once the same chip was used only for video games, but it has now taken a turn that brings it closer to working with AI.
To increase the speed of algorithm training, the remote supercomputer-led facility is expected to include 16,000 GPUs and 1 exabyte of storage, equal to one trillion gigabytes or one million terabytes, giving it the ability to deliver data at 16 terabytes per second. If completed soon, RSC will become the world’s fastest artificial intelligence supercomputer.
The fastest supercomputer in the world today
According to the Top500 organization and the latest supercomputer list, the world’s fastest supercomputer will be the Japanese Fugaku machine, which uses no GPU, which clocked a top speed of 442 petaflops, with permutation understood as a unit of measurement. power, a petaflop represents a billion operations per second; However, for AI applications specifically, Fugaku will reach speeds of more than an exaflop, which is one million trillion operations per second; However, it will be taken off by the RSC, which aims for 5 exaflops.
Meta’s competition in AI is not small, because in order to be the world’s fastest supercomputer, RSC must end soon, as other machines like Frontier, which is already under construction, are expected to be three times faster than Fugaku.
Now, the more supercomputers running, the more applications society can better see into the many routines of life in the coming years.