Capacity Planning for AI Training Hardware
summary
Meta describes mCP as a process that turns product forecasts into firm hardware orders using capacity contracts and a unified resource schema so new accelerators plug in without rewriting tooling. Planning targets include ~600k GPUs by end of 2024 across a multivendor fleet (NVIDIA, AMD, Meta silicon) and regional collocation to avoid cross‑region traffic. Decision engines run at different cadences: the mCP solver on a quarterly cycle, Tetris solver hourly/daily for data placement, and a real‑time job placer; Infra Cloud Portal and ML Hub expose the flows.
AI-written summary · about
chapters (8)
- 0:00Introduction to MCP
- 2:24Growth and complexity
- 4:15AI training challenges
- 7:16AI foundations
- 14:26Solver and runtime operations
- 16:01User experience interfaces
- 17:55Threads launch case study
- 20:13Future outlook and closing
related talks
23mSupporting AI holistically Across the Infra stack
19mBuilding Meta's Internal Cloud: the opportunities and challenges
09mAI for NI
23mKeynote
20mGlobal Capacity Management
21mExtending Meta to the Public Cloud
27mLive Panel: Building at the Roofline — Planning and Delivering Gigawatt-Scale AI Infrastructure
17mKeynote