Megatron-Turing NLG 530B
Megatron-Turing NLG 530B is an AI model developed by Microsoft and NVIDIA (United States), first published in October 2021. It works in the language domain, on tasks such as language modeling, language modeling/generation and question answering. It counts among the frontier models: the systems trained with the most compute of their moment.
Training it took an estimated 8.6×10²³ FLOP of compute (estimation method: third-party estimation). The model has 530,000,000,000 parameters. It was trained on roughly 270B datapoints. Training ran on 4,480 NVIDIA A100 SXM4 80 GB for about 770 hours. The compute alone is estimated at $4M in 2023 dollars.
Access: Unreleased. Its weights are not openly released. The reference paper has 847 citations. Epoch AI rates the confidence of this record as confident.
- Organization
- Microsoft, NVIDIA
- Country of organization
- United States
- Domain
- Language
- Task
- Language modeling, Language modeling/generation, Question answering
- Training compute
- 8.6×10²³ FLOP
- Compute estimation method
- Third-party estimation
- Parameters
- 530,000,000,000
- Dataset size
- 270B
- Training hardware
- NVIDIA A100 SXM4 80 GB
- Chips used
- 4,480
- Training time
- 770 h
- Chip-hours
- 3.4M
- Training power draw
- 3.6 MW
- Training cost (2023 USD)
- $4M
- Numerical format
- BF16
- Model accessibility
- Unreleased
- Open weights
- No
- Citations
- 847
- Epoch confidence
- Confident
| 01 | LAMBADA | Score | 0.87 |
| 02 | BoolQ | Score | 0.85 |
| 03 | PIQA | Score | 0.83 |
| 04 | HellaSwag | Overall accuracy | 0.82 |
| 05 | WinoGrande | Accuracy | 0.79 |
| 06 | Adversarial NLI | Score | 0.4 |