AWS Fast BERT Training

Compute

129 H100e

Chips

2,048

First operational

Dec 3, 2019

AWS Fast BERT Training is a large GPU cluster operated by Amazon, located in the United States, in the private sector. It came online around December 2019. At that moment it ranked number 11 among known clusters.

It is built on 2,048 NVIDIA V100. That amounts to about 129 H100e of AI compute. The cluster draws roughly 1.36 MW. Its hardware is estimated at $40M.

Epoch AI rates the certainty of this record as confirmed.

Full record

Owner: Amazon
Users: Amazon
Sector: Private
Status: Existing
Country: United States
Primary chip: NVIDIA V100
Primary chip count: 2,048
Total AI chips: 2,048
H100 equivalents: 129
Power capacity: 1.36 MW
Hardware cost: $40M
First operational: Dec 3, 2019
Rank when it came online: 11
Certainty: Confirmed