Transformer (Adaptive Input Embeddings) WT103

Training compute

4.5×10¹⁹ FLOP

Parameters

247M

Published

Sep 28, 2018

Transformer (Adaptive Input Embeddings) WT103 is an AI model developed by Facebook AI Research (United States and France), first published in September 2018. It works in the language domain, on tasks such as language modeling.

Training it took an estimated 4.5×10¹⁹ FLOP of compute (estimation method: hardware,operation counting). The model has 247,000,000 parameters. It was trained on roughly 100M datapoints. Training ran on 8 NVIDIA V100 for about 67 hours. The compute alone is estimated at $3K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 433 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Facebook AI Research
Country of organization: United States, France
Domain: Language
Task: Language modeling
Training compute: 4.5×10¹⁹ FLOP
Compute estimation method: Hardware, Operation counting
Parameters: 247,000,000
Dataset size: 100M
Training hardware: NVIDIA V100
Chips used: 8
Training time: 67 h
Chip-hours: 4.3K
Training power draw: 5.0 kW
Training cost (2023 USD): $3K
Numerical format: FP16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 433
Epoch confidence: Confident