ViT-22B

Training compute

1.9×10²³ FLOP

Parameters

21.7B

Published

Feb 10, 2023

ViT-22B is an AI model developed by Google (United States), first published in February 2023. It works in the vision domain, on tasks such as object detection and image classification.

Training it took an estimated 1.9×10²³ FLOP of compute (estimation method: hardware). The model has 21,743,000,000 parameters. It was trained on roughly 4B datapoints. Training ran on 1,024 Google TPU v4 for about 347 hours. The compute alone is estimated at $286K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 862 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Google
Country of organization: United States
Domain: Vision
Task: Object detection, Image classification
Training compute: 1.9×10²³ FLOP
Compute estimation method: Hardware
Parameters: 21,743,000,000
Dataset size: 4B
Training hardware: Google TPU v4
Chips used: 1,024
Training time: 347 h
Training power draw: 694.9 kW
Training cost (2023 USD): $286K
Numerical format: BF16
Model accessibility: Unreleased
Open weights: No
Citations: 862
Epoch confidence: Confident