Gato

Training compute

4×10²¹ FLOP

Parameters

1.2B

Published

May 12, 2022

Gato is an AI model developed by DeepMind (United Kingdom), first published in May 2022. It works in the multimodal, robotics, games and language domain, on tasks such as atari, image captioning, chat and robotic manipulation.

Training it took an estimated 4×10²¹ FLOP of compute (estimation method: hardware,operation counting). The model has 1,180,000,000 parameters. It was trained on roughly 524.3B datapoints. Training ran on 256 Google TPU v3 for about 96 hours. The compute alone is estimated at $4K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 1,064 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: DeepMind
Country of organization: United Kingdom
Domain: Multimodal, Robotics, Games, Language
Task: Atari, Image captioning, Chat, Robotic manipulation
Training compute: 4×10²¹ FLOP
Compute estimation method: Hardware, Operation counting
Parameters: 1,180,000,000
Dataset size: 524.3B
Training hardware: Google TPU v3
Chips used: 256
Training time: 96 h
Training power draw: 231.3 kW
Training cost (2023 USD): $4K
Model accessibility: Unreleased
Open weights: No
Citations: 1,064
Epoch confidence: Confident