Why it matters
AI infrastructure operators can see how Meta is planning for hardware failures at very large scale, including by reserving a substantial share of GPUs for redundancy.
What we know
Confirmed by several sources
- Meta published details about Prometheus, a supercluster in New Albany, Ohio, used for AI model training. — HPCwire, The Register
- Prometheus uses Nvidia Grace Blackwell systems, and Meta expects to add a Vera Rubin-based platform. — HPCwire, The Register
Reported by one source
- Prometheus was used to train Meta’s Muse family of AI models. — HPCwire
- More than 11 percent of the GPUs are dead or held idle as replacements for failures. — The Register
- Prometheus is expected to scale to a gigawatt of capacity and about a million accelerators. — The Register
- Meta’s Catalina cluster uses GB200 chips, an NVL72 architecture, and a two-rack configuration it calls Ariel, with 800G of I/O per GPU. — HPCwire
Coverage 2 publishers
The RegisterFirst report
More than 11 percent of the GPUs at Meta’s Prometheus supercluster in New Albany, Ohio, are either dead or sitting idle waiting to step in when the next drops out, Zuckercorp revealed this week. The massive pool of redundant GPUs is just…
-
Inside Prometheus, the Gigawatt Supercluster That Trained Meta’s Muse
Meta this week released details on Prometheus, the New Albany, Ohio-based supercluster that it used to train Muse, its latest family of AI model. The information is useful, if only to catch a glimpse of how HPC can evolve when price is no…
Earliest report first. Articles stay on their publishers’ sites; each link opens the original.