Training clusters
Drawn to scale
The machines frontier models are trained on: accelerators inside servers, servers inside racks, racks across a data hall. Pick one to see its layout, power draw, and how quickly it could have trained known models.
Colossus
xAI, United States (2024)
+1,323 more
- Accelerators
- 100,000
- Racks
- 1,563
- Per rack
- 64 chips
- Power
- 77.0 MW
- Chip
- H100 SXM (80GB)
- Peak
- 98.90 EFLOP/s
Eagle
Microsoft Azure, United States (2023)
+210 more
- Accelerators
- 14,400
- Racks
- 450
- Per rack
- 32 chips
- Power
- 11.1 MW
- Chip
- H100 SXM (80GB)
- Peak
- 14.24 EFLOP/s
Research SuperCluster (RSC)
Meta, United States (2022)
+260 more
- Accelerators
- 16,000
- Racks
- 500
- Per rack
- 32 chips
- Power
- 7.0 MW
- Chip
- A100 SXM (80GB)
- Peak
- 4.99 EFLOP/s
TPU v4 Pod (single pod)
Google, United States (2021)
- Accelerators
- 4,096
- Racks
- 64
- Per rack
- 64 chips
- Power
- 0.9 MW
- Chip
- TPU v4
- Peak
- 1.13 EFLOP/s
Looking for the estimator, validated training runs, and regulatory threshold comparisons? Open the full dashboard.