Turing-NLG

Training compute

1.6×10²² FLOP

Parameters

17B

Published

Feb 13, 2020

Turing-NLG is an AI model developed by Microsoft (United States), first published in February 2020. It works in the language domain, on tasks such as text autocompletion, language generation and text summarization.

Training it took an estimated 1.6×10²² FLOP of compute (estimation method: third-party estimation,operation counting). The model has 17,000,000,000 parameters. It was trained on roughly 46.4B datapoints. Training ran on 256 NVIDIA Tesla V100 DGXS 32 GB. The compute alone is estimated at $52K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 114 citations. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Microsoft
Country of organization: United States
Domain: Language
Task: Text autocompletion, Language generation, Text summarization
Training compute: 1.6×10²² FLOP
Compute estimation method: Third-party estimation, Operation counting
Parameters: 17,000,000,000
Dataset size: 46.4B
Training hardware: NVIDIA Tesla V100 DGXS 32 GB
Chips used: 256
Training power draw: 130.9 kW
Training cost (2023 USD): $52K
Model accessibility: Unreleased
Open weights: No
Citations: 114
Epoch confidence: Likely