The AI of tomorrow requires more powerful computers that can run quintillions of operations every second. These computers must be powerful enough to process massive datasets extremely fast and train large AI models. For more building and researching more advanced implementations of computer vision and Natural Language Processing (NLP), Meta has announced that they have built a new supercomputer. 

Meta, formerly Facebook, has always invested in improving or scaling its IT architecture, platform and systems. The company has stated that this powerful AI supercomputer will be one of the solutions that will help develop the metaverse. Meta’s new AI supercomputer is one of the amazing products of research in the fields of computing and Artificial Intelligence.

This supercomputer or the AI Research SuperCluster (RSC) as Meta has named it is reported to be one of the fastest AI supercomputers in the world today. As a matter of fact, according to Meta, the AI RSC will become the fastest AI supercomputer in the world once it is fully built by the middle of this year (2022).

Meta’s researchers have already started training large NLP and computer vision models to further develop the system. Meta’s target is to eventually train AI models that comprise trillions of parameters.

The metaverse is a virtual reality platform that is expected to eventually cater to millions if not billions of users online. The virtual reality experience and the digital spaces or possessions of these users must all be closely assisted by AI to provide a seamless experience. 

The AI RSC will be very useful in developing the metaverse and in also governing the experience of users. For instance, let us take an example of bullying in virtual reality where the miscreant is not saying anything or typing anything and simply carrying out some offending action. This can be easily handled with the help of advanced AI that can identify unwanted behaviour. 

What are Supercomputers?

Supercomputers can be defined as powerful high-speed computers that can carry out thousands of complex calculations in real-time. These computers have much more processing capacity and other powerful hardware resources that personal computers do not have. Even when compared with personal computers that have the highest specifications in terms of RAM and processing power, supercomputers are in a league of their own, far beyond the reach of personal or corporate computing.

A supercomputer like Meta’s would be even faster and more powerful than the traditional supercomputers that are used by government agencies. In 2019, Google’s AI division announced a quantum processor named Sycamore that could complete a mathematical task that would take the best supercomputer of that time (IBM’s Summit) 10,000 years to finish, in about 200 seconds.

If Meta’s AI supercomputer is more powerful than that, we cannot even begin to imagine how much capacity the system will have in terms of computing and processing.

Meta’s Supercomputer

Meta’s supercomputer has been developed to specifically train Machine Learning systems and build AI implementations that fall under augmented reality, computer vision, natural language processing, image detection and many more. The AI RSC will help researchers develop advanced AI models that can process and learn from trillions of examples. 

This AI supercomputer will be capable of working with hundreds of human languages and analysing millions of images, audio, videos and strings of text in real-time. Seamless augmented reality in the metaverse will become possible with a supercomputer as capable as the AI RSC. Meta’s developers are very proud of their creation and are ready to vouch for its powerful processing capacity. 

The supercomputer will be working with real-world data from Meta’s own systems and platforms, thus helping Meta’s researchers develop a more advanced AI as compared to AI that is trained only with publicly available datasets or open-source data. The RSC will be used for training a large range of systems and models for Meta such as content moderation algorithms that identify and take down hate speech. 

Augmented reality features will also be improved and developed in such a way that it provides a truly immersive experience for users. All of this will reinstate the company as a market leader in virtual spaces. The research for this high-performance system has been going on for more than a year and a half. The processing, cabling, power, cooling and networking for the supercomputer were all designed from scratch in-house by Meta’s engineers.

Powerful computing infrastructure is essential for AI supercomputers, and Meta’s AI research team has been developing high-performance systems for years now. The first instance of high-performance infrastructure by Meta was seen in 2017, and it boasted 22,000 NVIDIA V100 Tensor Core Graphics Processing Units in each cluster. Each of these could carry out 35,000 training processes daily. Due to the bar already being set quite high by Meta’s developers in terms of reliability and performance, the AI RSC is truly an engineering marvel. 

Developing the AI RSC

The first phase of the AI RSC is already built and is operating as desired. Phase one is built using 6,080 connected GPUs and consists of 760 Nvidia GGX A100 systems. This makes the AI supercomputer exceptionally great at handling machine learning problems. This version of the AI RSC is already 20 times more powerful than previous AI supercomputers. The AI RSC has created new benchmarks for other companies to follow suit and is providing a massive improvement in performance for models such as computer vision models.

The AI RSC’s phase two will be completed before the end of 2022, and at that point, it will contain 16,000 GPUs. With this capacity, the AI supercomputer will be capable of training AI systems that will be built on datasets that have trillions of parameters and are as large as an exabyte. For reference, 1 exabyte is 1000 petabytes and 1 petabyte is 1000000 gigabytes or GB (1000 TB).

Microsoft’s AI supercomputer, another extremely powerful supercomputer, contains only 10,000 GPUs. The AI RSC is being developed to become the most powerful AI supercomputer that has ever existed till now. Microsoft and Nvidia have also announced their own AI supercomputers but, based on Meta’s statements, they will probably not be as powerful as the AI RSC. 

Conclusion

The AI RSC will pave the way for future AI supercomputers and will help Meta build more advanced technologies. More than anything, the AI RSC will have a huge part to play in the development of the metaverse. The tremendous computing capabilities of this AI supercomputer will enable Meta to build more accurate AI models for years to come. 

Meta’s long-term investments in machine learning and AI computing have paid off, and this next generation of AI infrastructure will also broadly help the entire AI community progress.

Also Read: Meta Plans To Set Up An Ai-Powered Universal Speech Translator