Author: Jim O Brien/CEO

Developing the next generation of advanced AI will require powerful new computers capable of quintillions of operations per second. Today, Meta is announcing that they designed and built the AI Research SuperCluster (RSC) — which they believe is among the fastest AI  supercomputers running today and will be the fastest AI supercomputer in the world when it’s fully built out in mid-2022. Their researchers have already started using RSC to train large models in natural language processing (NLP) and computer vision for research, with the aim of one day training models with trillions of parameters. 

RSC will help Meta’s AI researchers build new and better AI models that can learn from trillions of examples; work across hundreds of different languages; seamlessly analyze text, images,  and video together; develop new augmented reality tools; and much more. Our researchers will  be able to train the largest models needed to develop advanced AI for computer vision, NLP, speech recognition, and more. We hope RSC will help us build entirely new AI systems that can,  for example, power real-time voice translations to large groups of people, each speaking a  different language, so they can seamlessly collaborate on a research project or play an AR  game together. Ultimately, the work done with RSC will pave the way toward building  technologies for the next major computing platform — the metaverse, where AI-driven  applications and products will play an important role.

Why do we need an AI supercomputer at this scale? 

Meta has been committed to long-term investment in AI since 2013, when we created the  Facebook AI Research lab. In recent years, we’ve made significant strides in AI thanks to our leadership in a number of areas, including self-supervised learning, where algorithms can learn  from vast numbers of unlabeled examples, and transformers, which allow AI models to reason  more effectively by focusing on certain areas of their input.  

To fully realize the benefits of self-supervised learning and transformer-based models, various  domains, whether vision, speech, language, or for critical use cases like identifying harmful  content, will require training increasingly large, complex, and adaptable models. Computer  vision, for example, needs to process larger, longer videos with higher data sampling rates.  Speech recognition needs to work well even in challenging scenarios with lots of background  noise, such as parties or concerts. NLP needs to understand more languages, dialects, and  accents. And advances in other areas, including robotics, embodied AI, and multimodal AI, will  help people accomplish useful tasks in the real world. 

High-performance computing infrastructure is a critical component in training such large models,  and Meta’s AI research team has been building these high-powered systems for many years.  The first generation of this infrastructure, designed in 2017, has 22,000 NVIDIA V100 Tensor  Core GPUs in a single cluster that performs 35,000 training jobs a day. Up until now, this  infrastructure has set the bar for Meta’s researchers in terms of its performance, reliability, and  productivity.  

In early 2020, we decided the best way to accelerate progress was to design a new computing infrastructure from a clean slate to take advantage of new GPU and network fabric technology. We wanted this infrastructure to be able to train models with more than a trillion parameters on  data sets as large as an exabyte — which, to provide a sense of scale, is the equivalent of   36,000 years of high-quality video.

While the high-performance computing community has been tackling scale for decades, we also  had to make sure we have all the needed security and privacy controls in place to protect any  training data we use. Unlike with our previous AI research infrastructure, which leveraged only open source and other publicly available data sets, RSC also helps us ensure that our research  translates effectively into practice by allowing us to include real-world examples from Meta’s  production systems in model training. By doing this, we can help advance research to perform  downstream tasks such as identifying harmful content on our platforms as well as research into embodied AI and multimodal AI to help improve user experiences on our family of apps. We  believe this is the first time performance, reliability, security, and privacy have been tackled at such a scale.  supercomputers are built by combining multiple GPUs into compute nodes, which are then  connected by a high-performance network fabric to allow fast communication between those  GPUs. RSC today comprises a total of 760 NVIDIA DGX A100 systems as its compute nodes,  for a total of 6,080 GPUs — with each A100 GPU being more powerful than the V100 used in  our previous system. The GPUs communicate via an NVIDIA Quantum 200 Gb/s InfiniBand  two-level Clos fabric that has no oversubscription. RSC’s storage tier has 175 petabytes of Pure  Storage FlashArray, 46 petabytes of cache storage in Penguin Computing Altus systems, and  10 petabytes of Pure Storage FlashBlade. 

Early benchmarks on RSC, compared with Meta’s legacy production and research  infrastructure, have shown that it runs computer vision workflows up to 20 times faster, runs the  NVIDIA Collective Communication Library (NCCL) more than nine times faster, and trains large scale NLP models three times faster. That means a model with tens of billions of parameters  can finish training in three weeks, compared with nine weeks before.

Designing and building something like RSC isn’t a matter of performance alone but performance  at the largest scale possible, with the most advanced technology available today. When RSC is  complete, the InfiniBand network fabric will connect 16,000 GPUs as endpoints, making it one of  the largest such networks deployed to date. Additionally, we designed a caching and storage  system that can serve 16 TB/s of training data, and we plan to scale it up to 1 exabyte.  

All this infrastructure must be extremely reliable, as we estimate some experiments could run for  weeks and require thousands of GPUs. Lastly, the entire experience of using RSC has to be  researcher-friendly so our teams can easily explore a wide range of AI models.

 A big part of achieving this was in working with a number of long-time partners, all of whom also helped design the first generation of our AI infrastructure in 2017. Penguin Computing, our  architecture and managed services partner, worked with our operations team on hardware  integration to deploy the cluster and helped set up major parts of the control plane. Pure  Storage provided us with a robust and scalable storage solution. And NVIDIA provided us with  its AI computing technologies featuring cutting-edge systems, GPUs, and InfiniBand fabric, and  software stack components like NCCL for the cluster. …and doing it remotely, during a pandemic 

But there were other unexpected challenges that arose in RSC’s development — namely the  coronavirus pandemic. RSC began as a completely remote project that the team took from a  simple shared document to a functioning cluster in about a year and a half. COVID-19 and  industry-wide wafer supply constraints also brought supply chain issues that made it difficult to  get everything from chips to components like optics and GPUs, and even construction materials  — all of which had to be transported in accordance with new safety protocols. To build this  cluster efficiently, we had to design it from scratch, creating many entirely new Meta-specific  conventions and rethinking previous ones along the way. We had to write new rules around our  data center designs — including their cooling, power, rack layout, cabling, and networking  (including a completely new control plane), among other important considerations. We had to  ensure that all the teams, from construction to hardware to software and AI, were working in  lockstep and in coordination with our partners. 

Beyond the core system itself, there was also a need for a powerful storage solution, one that  can serve terabytes of bandwidth from an exabyte-scale storage system. To serve AI training’s  growing bandwidth and capacity needs, we developed a storage service, AI Research Store  (AIRStore), from the ground up. To optimize for AI models, AIRStore utilizes a new data  preparation phase that preprocesses the data set to be used for training. Once the preparation  is performed one time, the prepared data set can be used for multiple training runs until it  expires. AIRStore also optimizes data transfers so that cross-region traffic on Meta’s inter datacenter backbone is minimized. 

How we safeguard data in RSC 

To build new AI models that benefit the people using our services — whether that’s detecting  harmful content or creating new AR experiences — we need to teach models using real-world  data from our production systems. RSC has been designed from the ground up with privacy and  security in mind, so that Meta’s researchers can safely train models using encrypted user generated data that is not decrypted until right before training. For example, RSC is isolated  from the larger internet, with no direct inbound or outbound connections, and traffic can flow  only from Meta’s production data centers. 

To meet our privacy and security requirements, the entire data path from our storage systems to  the GPUs is end-to-end encrypted and has the necessary tools and processes to verify that  

these requirements are met at all times. Before data is imported to RSC, it must go through a  privacy review process to confirm it has been correctly anonymized. The data is then encrypted  before it can be used to train AI models and decryption keys are deleted regularly to ensure  older data is not still accessible. And since the data is only decrypted at one endpoint, in  memory, it is safeguarded even in the unlikely event of a physical breach of the facility. 

Phase two and beyond 

RSC is up and running today, but its development is ongoing. Once we complete phase two of  building out RSC, we believe it will be the fastest AI supercomputer in the world, performing at  nearly 5 exaflops of mixed precision compute. Through 2022, we’ll work to increase the number  of GPUs from 6,080 to 16,000, which will increase AI training performance by more than 2.5x.  The InfiniBand fabric will expand to support 16,000 ports in a two-layer topology with no  oversubscription. The storage system will have a target delivery bandwidth of 16 TB/s and  exabyte-scale capacity to meet increased demand. 

We expect such a step function change in compute capability to enable us not only to create  more accurate AI models for our existing services, but also to enable completely new user  experiences, especially in the metaverse. Our long-term investments in self-supervised learning  and in building next-generation AI infrastructure with RSC are helping us create the foundational  technologies that will power the metaverse and advance the broader AI community as well.