Switch

Training compute

8.2×10²² FLOP

Parameters

1.6T

Published

Jan 11, 2021

Switch is an AI model developed by Google (United States), first published in January 2021. It works in the language domain, on tasks such as text autocompletion. It counts among the frontier models: the systems trained with the most compute of their moment.

Training it took an estimated 8.2×10²² FLOP of compute (estimation method: third-party estimation). The model has 1,571,000,000,000 parameters. It was trained on roughly 86.4B datapoints. Training ran on 1,024 Google TPU v3 for about 648 hours. The compute alone is estimated at $145K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 3,888 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Google
Country of organization: United States
Domain: Language
Task: Text autocompletion
Training compute: 8.2×10²² FLOP
Compute estimation method: Third-party estimation
Parameters: 1,571,000,000,000
Dataset size: 86.4B
Training hardware: Google TPU v3
Chips used: 1,024
Training time: 648 h
Chip-hours: 663.6K
Training power draw: 935.4 kW
Training cost (2023 USD): $145K
Numerical format: BF16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 3,888
Epoch confidence: Confident