A supercomputer by the end of 2022? Social media company Facebook made waves last year by re-branding itself as Meta. Now it’s turning heads with a new Artificial Intelligence (AI) supercomputer – currently in development, and scheduled for completion by mid-2022.
The Battle for Better AI
Computers are already pretty good at processing natural language, recognizing speech and identifying objects in images and videos – but there’s always room for improvement. The way to do that is by training the algorithms. Help them learn more, and learn better, by showing them more examples, tweaking the parameters and incorporating the adjustments. This training is now done by self-learning systems. Feed trillions of examples into the system, adjust the learning parameters, and AI systems get smarter on their own.
AI is the playing field for the digital giants today and every company is working hard to make strides and improve their services. It’s a resource-heavy task in terms of data, but that’s what all these companies have in spades. Meta generates data from multiple channels – Facebook, Instagram, WhatsApp, and Oculus – more than enough to start trying to eke out competitors like Microsoft, Google, Apple, and Nvidia, by outpacing them in the training arena. Meta believes its supercomputer will help the company process that data better, faster, and with less human involvement.
Supercharged Content and Business Intelligence
Meta’s “RSC” supercomputer has been in development for a year and a half already at the AI Research SuperCluster (RSC) (and that’s where it got its name). Work so far has been focused primarily on improvements in natural language processing, speech recognition, and computer vision to help Meta improve its existing business services. Better algorithms will improve content moderation, helping Meta to better detect harmful and inappropriate speech, imagery and videos to help ensure the security and safety of digital spaces.
Advances in AI will drive Meta’s plans down the road as well, helping drive its forthcoming Augmented Reality systems and hardware, plus aspects of Meta’s metaverse, the virtual experiences announced last year as the company’s new centerpiece. Improvements in natural language processing, for instance, could help deliver seamless voice translation in virtual office spaces, games, shops, and more.
Meta’s research blog describes what they’re aiming for with these leaps of learning: “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.”
According to company benchmarking, the RSC already outpaces Meta’s legacy systems. Computer vision workflows are performing 20 times faster than they did with older processing infrastructure, and the system is training large natural language processing models three times faster than its legacy systems.
The Nuts and Bolts
For the number-hungry folks out there who want to know how Meta’s RSC measures up: the supercomputer currently consists of 6,080 Nvidia A100 GPUs, linked through a Quantum InfiniBand network that can transmit data at 200 Gigabits per second. In terms of storage, we’re talking petabytes, each one equal to a million GB. The system can accommodate 175 petabytes (PB) of bulk storage, 46 PB of cache storage and 10 PB of NFS storage.
And that’s just the beginning. RSC is up and running, but there’s more development in store. A second development stage will add nearly 10,000 GPUs and increase training performance by 2.5X. the company believes that once Phase 2 is complete, the RSC will be the fastest AI supercomputer in the world.
Powering Virtual Worlds
This colossal increase in computing capability and self-supervised learning, Meta hopes, will not only bring measurable benefits to current services. Meta wants it to power and support all kinds of novel user experiences in the metaverse.
Meta’s investment in supercomputing reminds us that AI isn’t magic. It takes major investment, ingenuity and innovation to make the incremental improvements that make digital experiences seamless and intuitive – for now and in our virtual future. Meta’s visionary infrastructural commitment to processing power and algorithmic refinement shows it’s ready to power up AI – and us – to the next level. We may be meeting up in the metaverse sooner than we expect.