Eagle
Microsoft Azure, United States (2023) — 14,400 x NVIDIA H100 SXM (80GB)
8 × H100 SXM (80GB) per server
4 servers per rack
= 32 accelerators per rack, drawing 22.4 kW (accelerators only)
Rack layout Estimated
NVIDIA HGX/DGX H100 carries 8 H100s (8U, ~10 kW); 4 per rack is a typical air-cooled density.
450 racks holding 1,800 servers and 14,400 accelerators.
At roughly 3 m² per rack including aisle space, that's about 1,350 m² of floor — roughly 3.2 basketball courts. Floor area is estimated: real halls also carry cooling, power, and networking equipment this doesn't count.
If fully dedicated, this cluster could train...
| Known training run | Time on this cluster | Originally took |
|---|---|---|
| GPT-3 (175B) | 15.3 hours | duration not published |
| Llama 2 (70B) | 1.7 days | 35.0 days on 2,048 x A100 SXM (80GB) |
| Llama 3.1 (405B) | 77.2 days | 54.0 days on 16,384 x H100 SXM (80GB) |
| PaLM (540B) | 5.2 days | duration not published |
| Chinchilla (70B) | 1.2 days | duration not published |
Hypothetical: assumes the whole cluster runs one job at the vendor-default MFU, with no downtime, restarts, or contention — Estimated. "Originally took" is the run's own reported hardware and duration where published, for contrast.
Source: TOP500 list (June 2023): Microsoft's Eagle supercomputer ranked #3 globally with 14,400 NVIDIA H100 GPUs.