Meta GenAI RoCE cluster
Meta uses this internal cluster to train Llama and other generative AI models.
Meta GenAI RoCE cluster is an operational GPU cluster run by Meta. It has 24,576 NVIDIA H100 GPUs installed.
Facts
Source per value- Owner / operator Organization that owns or operates the GPUs. Hosting and facility roles are listed under Partners.
- Meta 1
- GPU maker NVIDIA, AMD or Intel, taken from the GPU model.
- NVIDIA 1
- GPU model Accelerator model or family as disclosed. A GH200 is one GPU on a superchip; an MI300A is an APU that contains a GPU.
- NVIDIA H100 1
- GPUs installed Physical GPUs in the cluster, from the operator or a deterministic sum of disclosed quantities (nodes × GPUs per node). Planned counts, H100-equivalents and unaudited press totals are never counted here.
- 24,576 1
- GPU count notes What the count covers (one partition, one site, a design limit), qualifiers and competing figures.
- One of two separate physical clusters; not Meta fleet total. 1
- Location City, region and country where the cluster sits. Empty when the operator does not name the site.
- Not published
- Live power MW Power in service, in megawatts of IT load where the source says so. Company figures are marked in the notes; Epoch AI satellite estimates are labelled as estimates.
- Not published
- Status operational = running and named as such by the operator or a primary source; partly operational = a phase is live and more is coming; under construction = building underway; announced = disclosed or contracted with no confirmed go-live.
- operational 1
- Status notes The source’s wording on status, with the date.
- In use for Llama 3 training in March 2024 1
- Interconnect Cluster network fabric.
- 400 Gbps RoCE Ethernet 1
Sources
- engineering.fb.com accessed 2026-11-11
Related clusters
All 87 clustersLists
Need GPU capacity at this scale? Request quotes from GPU cloud providers
Tell us what you need once. We send it to matching GPU cloud providers and they reply to you directly.