Meta Platforms is ready to share the details of its new AI Research SuperCluster (RSC). In the joint press release with Nvidia, the company released the information about the new supercomputer which would be constructed by mid-2022. This new technology would be the world’s most powerful and fastest supercomputer. It would also be the fastest AI system in the world as claimed by Meta CEO Mark Zuckerberg.
Nvidia, which is the GPU designer for the supercomputer, will power Meta’s RSC. Nvidia has made a name for itself in artificial intelligence hardware technology. This is not the first time they are partnering with Meta, previously known as Facebook. in 2017, the Facebook AI Research lab made a supercomputer with around 22,000 Nividia v100 Tensor Core GPUs. The tech manages to perform around 35000 training jobs in just one day. Until now, this system worked as Meta’s main AI supercomputer.
The new supercomputer’s development was delayed by the COVID-19 pandemic. The remote working and supply chain constraints pushed back the work by a significant amount of time. The development of this new supercomputer should have begun in 2020. The point of the new development is to increase the computing power and handle more advanced AI workloads. The new facility will have the 760 NVIDIA DGX A100 systems. This has been specifically been designed for high-end computing. The new system managed to cram around 6,080 GPUs in just one supercluster.
The new system will run everything around 20 times faster. Compared to the work that the previous supercomputer could do in about nine weeks, Meta confirmed that the new facility will be able to do in just three weeks. While the supercomputer is still under construction, the company expects it to connect around 16,00 GPU’s with endpoints. The caching and storage system would easily be able to hold 16 TB/s of training data.
Each system would have around eight A200 GPUs. By the time it is fully developed, the AI RSC will have the largest installation of Nvidia DGX A100 systems. They are connected by 200 GB/s of InfiniBand fabric, which would make it one of the largest networks. The storage of this supercomputer is extremely impressive, with around 175 petabytes of Pure Storage. According to Meta, once the cluster is all done, the Ai RSC will be able to perform 5 exaflops or rather, 5 quintillion calculations per second. the impressive numbers can easily display well the whole RSC is expected to work and what is it capable of.
The point of making these supercomputers is to make the dream of “Metaverse” come true. Meta is essentially working towards making a 3-D virtual environment for public gatherings. This Metaverse would also allow translating voices in real-time. Thus, people from different demographics could come together to interact and perform any range of activities.
The supercluster should also help Meta make better new AI systems. By developing new AI which can analyze texts, images, and videos all at the same time. These new user experiences are to be fueled with the help of AI RSC.
Meta is not the only one working on creating AI, as Intel has been known to increase its efforts to acquire an AI system. However, if everything goes to plan, the Meta supercomputer will be the largest and strongest in the world.