Nemotron-4 340B

Training compute

1.8×10²⁵ FLOP

Parameters

340B

Published

Jun 14, 2024

Nemotron-4 340B is an AI model developed by NVIDIA (United States), first published in June 2024. It works in the language domain, on tasks such as language modeling/generation, chat and question answering. It counts among the frontier models: the systems trained with the most compute of their moment.

Training it took an estimated 1.8×10²⁵ FLOP of compute (estimation method: operation counting,hardware). The model has 340,000,000,000 parameters. It was trained on roughly 9T datapoints. Training ran on 6,144 NVIDIA H100 SXM5 80GB for about 2.2K hours. The compute alone is estimated at $21M in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. Epoch AI rates the confidence of this record as confident.

Full record

Organization: NVIDIA
Country of organization: United States
Domain: Language
Task: Language modeling/generation, Chat, Question answering
Training compute: 1.8×10²⁵ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 340,000,000,000
Dataset size: 9T
Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 6,144
Training time: 2,200 h
Training power draw: 8.5 MW
Training cost (2023 USD): $21M
Numerical format: BF16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Confident