Gemini 1.0 Ultra training cluster A

Compute

4K H100e

Chips

28,672

First operational

Aug 20, 2023

Gemini 1.0 Ultra training cluster A is a large GPU cluster operated by Google, located in the United States, in the private sector. It came online around August 2023. At that moment it ranked number 10 among known clusters.

It is built on 28,672 Google TPU v4. That amounts to about 4K H100e of AI compute. The cluster draws roughly 20.23 MW. Its hardware is estimated at $260M.

Epoch AI rates the certainty of this record as likely. Believed to be trained on two sets of 7 pods of 4096 TPUs

Full record

Owner: Google
Users: Google
Sector: Private
Status: Existing
Country: United States
Primary chip: Google TPU v4
Primary chip count: 28,672
Total AI chips: 28,672
H100 equivalents: 4K
8-bit throughput: 7.9×10¹⁸ OP/s
Power capacity: 20.23 MW
Hardware cost: $260M
First operational: Aug 20, 2023
Rank when it came online: 10
Certainty: Likely