Meta will have the world’s fastest artificial intelligence (AI) supercomputer by the middle of this year. This is important to realizing the vision of Facebook’s parent company in the metaverse.
The supercomputer Meta invests in is called the AI Research SuperCluster (RSC) in its entirety. According to a blog post from Meta, it’s already among the fastest of its kind and would be the best when it’s fully expanded in the middle of this year.
- Meta is investing heavily in an artificial intelligence supercomputer, which should be the fastest in the world by the middle of this year.
- The computer is needed by Facebook’s parent company for advanced AI research, which is used, among other things, for real-time translations and moderation in social media and soon in the metaverse.
- Meta says that it is taking the necessary steps to anonymize and encrypt the training data.
It then acts against “nearly five exaflops of mixed-precision computation.” One exaflop is equal to one trillion (trillion trillion) calculations per second. Mixed-precision computing is about using different levels of precision in the same computing operation.
“The experiences we build for the metaverse require massive computing power of trillions of operations per second. The RSC will enable new AI models that can learn from billions of examples, understand hundreds of languages, and more,” CEO Mark Zuckerberg said in a Facebook post.
The experiences we build for the metaverse require massive computing power of trillions of operations per second.
Meta is hard at work on the metaverse, a next phase of the Internet in which interconnected online worlds and virtual and augmented reality play a major role. Products and applications that rely on AI become even more important in that context.
New AI systems enable real-time voice translations for large groups of people speaking many different languages so they can collaborate on a research project or play a game in augmented reality.
‘With the RSC, we can train models that understand the context of a message. It is about signals such as language, images and tone,’ explains a spokesperson for Meta. ‘Take a pointing finger, for example. It can mean something different in different cultures, but the meaning also depends on the context. An AI model can learn from signals to determine whether an action, sound, or image is bad or good. This research is necessary, among other things, to improve security in the metaverse in the future.’
Introducing the AI Research Supercluster
In 2020, it was decided to design a new computing infrastructure to take advantage of new technologies. It should be possible to train models with over a billion parameters on datasets that can be up to an exabyte in size. An exabyte is one billion gigabytes, the equivalent of 36,000 years of high-quality video.
Meta realizes that the world watches with great suspicion how you handle data. By the way, much more data will be collected in the metaverse than is already the case. The company explains that before entering the data in the RSC, it is verified that the data has been correctly anonymized. The data is then encrypted before it can be used to train AI models.
AI can help make the metaverse safer.
AI supercomputers are being built to combine graphics processing units (GPUs) into nodes, which are then interconnected in a network that enables lightning-fast communication between nodes. GPUs can process a large amount of data at once, which makes them particularly useful for AI, video editing, and gaming. Meta makes use of technology from US chip specialist Nvidia.
The RSC now uses 6080 GPUs and can train a model with tens of billions of parameters in three weeks compared to the previous nine weeks. The full RSC will have 16,000 GPUs.