Meta

Meta is the parent company of Facebook, but it also does a lot in AI. This page exists as a locus of backlinks for pages that touch on Meta’s contributions to AI and HPC.

As of November 2024, Meta has disclosed the following about its GPU infrastructure:

1. They use 50% of its GPU infrastructure on recommendation models.¹
2. They have 350K H100 GPUs²
3. They have 600K “H100 equivalents overall”²
   • Meta claims 48K H100s (Meta’s H100 clusters) is 200 EFLOPS FP8 at 2:1 sparsity, or 100 EFLOPS FP8
   • However, H100 is 1978.9 TFLOPS FP8 (not sparse), and knowing the exact number of GPUs in Meta’s H100 clusters, this true number is 97.267 EFLOPS peak at FP8
   • the difference of 250K H100-equivalents is therefore 494.725 EFLOPS FP8 (dense)

Footnotes

1. Balaji, Architectural Challenges in Modern AI Workloads. Workshop on Co-design of Next-Generation HPC Systems for Artificial Intelligence and Mixed Analytics. SC’24. ↩

2. Balaji, Herding Llamas: A Sneak Peek Into Meta’s Infrastructure for Generative AI. SC’24. ↩ ↩²