Meta Platforms Why the Facebook parent is building a giant AI computer
Californians are investing heavily in their "Metaverse" idea for the future.
Dusseldorf It should be able to understand hundreds of languages and process trillions of pieces of data: By summer, Meta Platforms, parent company of Facebook, Instagram and WhatsApp, plans to build the world’s most powerful computer for AI applications. The company made the announcement on Tuesday night.
Meta is silent on the location or cost of the AI Research Supercluster (RSC). To that end, the company announced impressive technical details. The supercomputer will have about 16.000 graphics processors from Nvidia, 175 petabytes of memory and a network speed of 200 gigabits per second. The speed for certain computational processes: nearly five exaflops, ten times faster than the most powerful supercomputer to date.
With the computer, Facebook hopes to advance its own research into artificial intelligence (AI). The goal is to develop "entirely new AI systems that, for example, enable real-time translation for large groups of people who all speak different languages," according to a blog post by the group. The technology could allow those people to collaborate on a research project or play a game together without communication problems, it said.
The computer is expected to pay off on the group’s central project for the future. "The Metaverse and its applications require enormous computing capacity," Meta CEO Mark Zuckerberg said. In the so-called metaverse, the virtual and real worlds will mix.
Find the best jobs now and
Be notified by email.
Back in October, Facebook had declared that it was working on an "ego AI" that would be trained with videos from the first-person perspective. If AI learns in the same way humans do, technology chief Mike Schroepfer wrote at the time, it could be used to create "smarter and more useful technologies. "Imagine an AI assistant that helps you remember where you left your keys or teaches you how to play the guitar."
Meta raises the bar for innovation
But experts say it will take several years for the new computer to make such experiences possible. "Just as a stonemason needs a chisel," says Chirag Dekate, analyst and computer expert at the consultancy Gartner, "RSC will first be a tool for meta researchers to design new AI models."
They’ll have their work cut out for them. Meta, for example, is working on an AI that pays attention not only to spoken language, but also to body posture, lip movement and other factors at the same time – to enable real-time, accurate translations into other languages. "With the computer, Meta is raising the bar for innovation," says Dekate.
Some applications are much closer: checking the posts and comments on Facebook or Instagram. Similar to a human, the new supercomputer should be able to build meaningful context from multiple input sources such as speech and images.
That would be a big help to the company, which has come under heavy criticism in recent months. Whistleblower Frances Haugen had made public how little Meta can verify its content. Here’s how the AI kept mistaking car accidents for cockfights, to the amazement of employees, the Wall Street Journal reported. According to the whistleblower’s documents, the AI detects only three to five percent of inflammatory calls and 0.6 percent of violence and riots.
But this task is central to Facebook’s business model; the quality of the network and the trust of users are essential. In addition, there are the costs. According to the internal documents, Facebook spent more than $100 million in 2019 to review the content through employees and surveys.
Artificial intelligence is greatly changing the demands on computers. Having to do many calculations at once, but less accurately than before – for example, to correctly classify an object in a photo, but not details such as color or the like.
The company has recognized the importance of hardware
Graphics processors are excellently suited for this purpose. A "Graphic Processing Unit" (GPU) is otherwise used for video games, such as to put as many pixels per second on the screen as possible.
The fact that Meta is developing its own supercomputer probably also has another background. The corporation has learned from the past that it is dangerous to rely only on software development in the competition of the most powerful corporations in the world.
Meta therefore also uses a lot of resources to develop its own glasses for VR (virtual reality) and AR (augmented resp. Augmented reality) to develop. The big challenge at the moment is to make the devices smaller, lighter and therefore suitable for everyday use.