DeBERTa

Training compute

2.6×10²² FLOP

Parameters

1.5B

Published

Jun 10, 2021

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

Training it took an estimated 2.6×10²² FLOP of compute (estimation method: hardware). The model has 1,500,000,000 parameters. It was trained on roughly 20.8B datapoints. Training ran on 256 NVIDIA V100 for about 720 hours. The compute alone is estimated at $7K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 3,764 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Microsoft
Country of organization: United States
Domain: Language
Task: Language modeling/generation, Question answering
Training compute: 2.6×10²² FLOP
Compute estimation method: Hardware
Parameters: 1,500,000,000
Dataset size: 20.8B
Training hardware: NVIDIA V100
Chips used: 256
Training time: 720 h
Training power draw: 155.4 kW
Training cost (2023 USD): $7K
Numerical format: FP16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 3,764
Epoch confidence: Confident