Megatron-BERT
Megatron-BERT is an AI model developed by NVIDIA (United States), first published in September 2019. It works in the language domain, on tasks such as language modeling/generation. It counts among the frontier models: the systems trained with the most compute of their moment.
Training it took an estimated 2.2×10²² FLOP of compute (estimation method: operation counting,third-party estimation). The model has 3,900,000,000 parameters. It was trained on roughly 7B datapoints. Training ran on 512 NVIDIA Tesla V100S PCIe 32 GB for about 374 hours. The compute alone is estimated at $172K in 2023 dollars.
Access: Unreleased. Its weights are not openly released. The reference paper has 2,766 citations. Epoch AI rates the confidence of this record as confident.
- Organization
- NVIDIA
- Country of organization
- United States
- Domain
- Language
- Task
- Language modeling/generation
- Training compute
- 2.2×10²² FLOP
- Compute estimation method
- Operation counting, Third-party estimation
- Parameters
- 3,900,000,000
- Dataset size
- 7B
- Training hardware
- NVIDIA Tesla V100S PCIe 32 GB
- Chips used
- 512
- Training time
- 374 h
- Chip-hours
- 191.5K
- Training power draw
- 262.6 kW
- Training cost (2023 USD)
- $172K
- Numerical format
- FP16
- Model accessibility
- Unreleased
- Open weights
- No
- Citations
- 2,766
- Epoch confidence
- Confident