Megatron-Turing NLG 530B
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- Microsoft,NVIDIA
- Organisation type
- Industry,Industry
- Country
- United States of America
- Published
- 11 October 2021
- Authors
- Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, Elton Zhang, Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi, Yuxiong He, Michael Houston, Saurabh Tiwary, Bryan Catanzaro
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling, Language modeling/generation, Question answering
- Approach
- Self-supervised learning
- Numerical format
- BF16
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 530B
- Training data
- 270,000,000,000 tokens
- Batch size
- 3,932,160
"Our training dataset consists of 339 billion tokens and we trained MT-NLG on 270 billions tokens by blending the 15 training datasets as described above. We also set aside 2% of our data for validation." 1 token ~ 0.75 words
"The sequence length is 2048 and the global batch size is 1920. We used 8-way tensor and 35-way pipeline parallelism. The learning rate is 5.0e −5 . We used one billion tokens for linear learning rate warmup. We used cosine decay for the learning rate targeting to reach 10% of its value over 340 billion tokens. Over the first 12 billion tokens, we started at a batch size of 32 and gradually increased the batch size in increments of 32, until we reach the final batch size of 1920" Final batch s…
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 8.6 × 10²³ FLOP
- How it was established
- Third-party estimation
https://www.lesswrong.com/posts/bGuMrzhJdENCo8BxX/nvidia-and-microsoft-releases-530b-parameter-transformer?commentId=HSJSNspKp94tFcSCx source: https://lair.lighton.ai/akronomicon/ 9938 PF-days * 3600 * 24 * 10^15 = 8.586432e+23 6ND estimate: 6 * 530B * 270B = 8.586000e+23
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA A100 SXM4 80 GB
- Chips used
- 4,480
- Chip-hours
- 3,449,600
- Wall-clock time
- 770 hours (32.1 days)
- Hardware utilisation
- HFU 30.2%
- Power draw
- 3.6 MW
- Compute cost
- $3,848,506
Total compute was 1.17*10^24 FLOP. They don't directly report the utilization and training speed when using the full Selene supercomputer with 560 DGX * 8 A100/DGX = 4480 GPUs. See section 2.3 Hardware Setup. At 280 DGX, the utilization is 126/312 = 40% and a batch takes 60 seconds; at 350, it is 39% for 50 seconds; at 420, it is 36% for 44 seconds. The overall utilization was 30.2% and the full cluster has 560 DGX. Dividing the total compute by the total performance of 4480 A100 at 30.2% util…
They don't directly report the utilization and training speed when using the full Selene supercomputer with 560 DGX * 8 A100/DGX = 4480 GPUs. See section 2.3 Hardware Setup. At 280 DGX, the utilization is 126/312 = 40% and a batch takes 60 seconds; at 350, it is 39% for 50 seconds; at 420, it is 36% for 44 seconds. The overall utilization was 30.2% and the full cluster has 560 DGX. HFU = 0.3020
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- Unreleased
- Training code
- Unreleased
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Frontier model
- Yes
- Foundation model
- Yes
- Likely above 10²³ FLOP
- Yes
- Why it is tracked
- SOTA improvement,Training cost
- Record confidence
- Confident
- Citations
- 847
The 105-layer, transformer-based MT-NLG improved upon the prior state-of-the-art models in zero-, one-, and few-shot settings
Sources
Where this record came from and when it was last checked.
- Reference
- Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model
- Last updated
- 25 May 2026
What the numbers mean
Where it came from
Megatron-Turing NLG 530B was published by Microsoft,NVIDIA, in United States of America, in October 2021. The organisation is categorised as industry,Industry.
It works in Language, and is recorded as doing language modeling, Language modeling/generation, Question answering.
Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.
Training and provenance
Training it took roughly 8.6 × 10²³ FLOP of computation, on NVIDIA A100 SXM4 80 GB — a measure of what producing the model cost, not of how fast it answers.
It was trained on about 270,000,000,000 tokens of text.
The reason it appears in this catalogue at all is sOTA improvement,Training cost.
Answers
Megatron-Turing NLG 530B — common questions
Who created Megatron-Turing NLG 530B?
Megatron-Turing NLG 530B was published by Microsoft,NVIDIA, based in United States of America, categorised as industry,Industry.
When was Megatron-Turing NLG 530B released?
Megatron-Turing NLG 530B was published in October 2021. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
What is Megatron-Turing NLG 530B used for?
Megatron-Turing NLG 530B works in Language, and is recorded as handling language modeling, Language modeling/generation, Question answering. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
How much compute was used to train Megatron-Turing NLG 530B?
Around 8.6 × 10²³ FLOP, on NVIDIA A100 SXM4 80 GB. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
What GPU do I need to run Megatron-Turing NLG 530B?
None. Megatron-Turing NLG 530B is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Is Megatron-Turing NLG 530B open source?
No. Megatron-Turing NLG 530B has not had its weights published, so it exists only as a service controlled by its owner.
How many parameters does Megatron-Turing NLG 530B have?
Megatron-Turing NLG 530B has 530B parameters. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.