GLaM

Training compute

3.6×10²³ FLOP

Parameters

1.2T

Published

Dec 13, 2021

GLaM is an AI model developed by Google (United States), first published in December 2021. It works in the language domain, on tasks such as language modeling/generation and question answering.

Training it took an estimated 3.6×10²³ FLOP of compute (estimation method: operation counting,hardware). The model has 1,200,000,000,000 parameters. It was trained on roughly 600B datapoints. Training ran on 1,024 Google TPU v4 for about 1.4K hours. The compute alone is estimated at $541K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 1,198 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Google
Country of organization: United States
Domain: Language
Task: Language modeling/generation, Question answering
Training compute: 3.6×10²³ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 1,200,000,000,000
Dataset size: 600B
Training hardware: Google TPU v4
Chips used: 1,024
Training time: 1,366 h
Chip-hours: 1.4M
Training power draw: 701.5 kW
Training cost (2023 USD): $541K
Numerical format: BF16
Model accessibility: Unreleased
Open weights: No
Citations: 1,198
Epoch confidence: Confident