/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Meta unveils AI Research SuperCluster, a supercomputer for ML training, claiming it will have 16K GPUs and be the world's fastest upon completion in mid-2022

Designed to train the next-generation of machine learning systems  —  Social media conglomerate Meta is the latest tech company to build an …

The Verge James Vincent

Context & Ripple Effects

Before Meta’s announcement, Nvidia and NERSC had positioned Perlmutter’s 6,144 A100 GPUs as the leading AI-focused supercomputer. Meta’s 16,000-GPU target escalates the scale benchmark from a research-lab system to an internal platform for a major consumer internet company.

The cluster became an early step in a longer compute buildout: Meta later paired RSC with plans for MTIA and video-processing chips and said it was training Llama 4 on a cluster of more than 100,000 H100 chips.

First-order effects

  • Meta’s research teams gain a planned dedicated training system sized well beyond the GPU count cited for Perlmutter, concentrating more ML experimentation inside the company.
  • Meta directly challenges Nvidia and NERSC’s AI-supercomputer performance benchmark with a larger planned GPU cluster.

Second-order effects

  • RSC gives Meta a platform on which to justify custom silicon: its later MTIA program targets both AI training and model execution rather than leaving the compute roadmap solely to general-purpose GPUs.
  • The comparison shifts competition from isolated model work toward the size and performance of the training infrastructure supporting it, raising the scale bar for other AI developers.

Third-order effects

  • Meta’s progression from RSC to custom accelerators and a 100,000-plus-H100 training cluster points to AI leadership becoming increasingly tied to an integrated stack of data-center capacity, accelerators, and models.
  • If this pattern persists across large AI developers, frontier-model development will favor organizations able to sustain repeated infrastructure expansions rather than one-off supercomputer deployments.

The trend: AI training is evolving from standalone GPU clusters into continuously expanded, vertically integrated compute stacks tied to proprietary models and chips.

Discussion

  • @nxthompson @nxthompson on x
    Meta has access to massive amounts of data. And now it has built the largest supercomputer in the world to train its AI. So many of the most important advances in AI in the future will come at a few giant companies. https://www.wsj.com/...
  • @camirusso Camila Russo on x
    facebook @Meta builds a supercomputer in private to better monetize users' data in its metaverse. ethereum builds a world computer in the open that allows users to control their data in the metaverse. not the same. https://twitter.com/...
  • @thedropnft @thedropnft on x
    The AI system will “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.” https://ai.facebook.com/...
  • @jackclarksf Jack Clark on x
    One of the reasons I massively care about building public AI infrastructure like the NAIRR is that otherwise the private sector is going to be able to ‘out-think’ the public sector/commons by virtue of having bigger and better computers. We're sleepwalking into end of democracy.
  • @ldignan Larry Dignan on x
    Zuck on Meta's RSC supercomputer: The experiences we're building for the metaverse require enormous compute power (quintillions of operations / second!)" https://www.techmeme.com/...
  • @gmseabra Gustavo de M. Seabra on x
    This looks so much beyond anything we can do at the University level, that if feels like we're playing with toy cars while they drive a Ferrari. Hopefully they'll have better algorithms... https://twitter.com/...
  • @nvidiahpcdev @nvidiahpcdev on x
    It took only 18 months from the initial idea to a working supercomputer for @Meta to build their new #AI Research SuperCluster! 760 NVIDIA DGX A100s w/ 6,080 A100 GPUs that delivers 1,895 PFLOPS of performance. Here's a video with more info: https://www.facebook.com/...
  • @nvidiaai Nvidia Ai on x
    .@Meta's AI supercomputer - the largest NVIDIA DGX A100 customer system to date - will deliver 5 exaflops of #AI performance with InfiniBand fabric and software enabling optimization across thousands of GPUs. https://blogs.nvidia.com/...
  • @sunset_city_rpg @sunset_city_rpg on x
    Hey @Meta stop stealing stuff from the evil megacorporation in my game. First you took the name, now you're taking the business model too? That's not a great look. #cyberpunk
  • @jsotterbach @jsotterbach on x
    This is a sheer massive scale! I'm very excited to see what this amount of compute will produce, but also am sad to see the widening of the capability gap. Hopefully some of the results can be amortized by smaller players. But even a “meager” 8 A100s is unobtainable for many ... …
  • @ericjhonsa Eric Jhonsa on x
    $FB built a supercomputer with 6,080 $NVDA A100 GPUs, with plans to have 16K. Also appear to be using a lot of Mellanox InfiniBand gear. No word on the CPU supplier, but good chance it's $AMD, given the November announcement and their HPC traction. https://ai.facebook.com/... htt…
  • @teestarnes Trevor Starnes on x
    Today is an exciting day for us at Pure! Today Meta announced their AI Research Super Cluster (RSC) and specifically, their validation of Pure as a strategic partner. Modern data pipelines and modern analytics are enabling organiza...https://www.linkedin.com/ ... https://ai.faceb…
  • @jackclarksf Jack Clark on x
    For perspective, at ~6,000 A100s, Facebook's newly announced AI cluster is on par with Perlmutter, the world's fifth fastest supercomputer (~6,000 A100s). Ultimately, FB is going to scale to ~16,000 A100s. https://ai.facebook.com/...
  • @downingark Frank Downing on x
    For reference, this would be nearly 3x the size of Tesla's current AI training supercomputer Both systems use Nvidia A100 GPUs and Infiniband networking https://ai.facebook.com/... https://blogs.nvidia.com/...
  • @schrep Mike Schroepfer on x
    It took some amazing system design to make this possible - the caching / storage can serve up training data at 16TB/s (!!). By summer there'll be 16,000 GPUs running with no oversubscription. A beast! https://ai.facebook.com/...
  • @metaai @metaai on x
    Meta is announcing the AI Research SuperCluster (RSC), our latest AI supercomputer 💻 for AI research. RSC will allow our researchers to do new, groundbreaking experiments in #AI. Learn more about RSC and the important role it will play: https://ow.ly/... https://twitter.com/...
  • @jjvincent James Vincent on x
    Microsoft has one, Nvidia has one, now Meta has one too: an AI supercomputer. Come for the news, stay for the discussion of the accuracy of different floating point operations (and why that differentiates an AI supercomputer and a regularsupercomputer) https://www.theverge.com/..…