PaLI

Training compute

1.7×10²³ FLOP

Parameters

16.9B

Published

Sep 14, 2022

PaLI is an AI model developed by Google (United States), first published in September 2022. It works in the language, vision and multimodal domain, on tasks such as visual question answering, language modeling/generation and image captioning.

Training it took an estimated 1.7×10²³ FLOP of compute (estimation method: operation counting,hardware). The model has 16,900,000,000 parameters. It was trained on roughly 143.5B datapoints. Training ran on 1,024 Google TPU v4 for about 240 hours. The compute alone is estimated at $51K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 983 citations. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Google
Country of organization: United States
Domain: Language, Vision, Multimodal
Task: Visual question answering, Language modeling/generation, Image captioning
Training compute: 1.7×10²³ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 16,900,000,000
Dataset size: 143.5B
Training hardware: Google TPU v4
Chips used: 1,024
Training time: 240 h
Chip-hours: 172K
Training power draw: 697.2 kW
Training cost (2023 USD): $51K
Model accessibility: Unreleased
Open weights: No
Citations: 983
Epoch confidence: Likely