Megatron-BERT

Closed weights NVIDIA 3.9B parameters September 2019

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
NVIDIA
Organisation type
Industry
Country
United States of America
Published
17 September 2019
Authors
Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, 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/generation
Approach
Self-supervised learning
Numerical format
FP16

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
3.9B

Table 4

Training data
6,960,000,000 tokens

"The resulting aggregate corpus contains 174 GB of deduplicated text." 174e9 bytes * (1 word / 5 bytes) * (4 tokens / 3 words) = 4.64e10 tokens

Batch size
524,288

"we set the batch size to 1024 and use a learning rate of 1.0e4 warmed up over 10,000 iterations and decayed linearly over 2 million iterations. Other training parameters are kept the same as (Devlin et al., 2018)." in Devlin et al (BERT), sequences are 512 tokens

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
2.2 × 10²² FLOP

A third-party source: https://lair.lighton.ai/akronomicon/ claims 5.7e22 The authors report experimenting on 1 V100 GPU and achieving throughput of 39 TFLOPS which is 30% of the peak throughput. Therefore the GPU has a peak throughput of 130 TFLOPS so it is specifically the NVIDIA V100S PCIe. https://images.nvidia.com/content/technologies/volta/pdf/volta-v100-datasheet-update-us-1165301-r5.pdf Param-based calculation: 6ND = 6*3.9e9*(2e6+1e4)*1024*512 = 2.5e22 FLOP 1024 is the batch size, 512 …

How it was established
Operation counting,Third-party estimation

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 Tesla V100S PCIe 32 GB
Chips used
512
Chip-hours
191,488
Wall-clock time
374 hours (15.6 days)

The 8.3B GPT-like arch took 2.1 days per epoch on 512 GPUs, batch size 512, sequence length 1024. An epoch was 68.5k iterations. BERT: batch size 1024, sequence length 512, 2e6 iterations total. Halving the model size should ~halve the iteration time. Doubling the batch size should ~double the iteration time. Halving the sequence length should ~quarter the iteration time (quadratic scaling). Hence 3.1e-5 days/iteration * 2 * 1/2 * 1/4 = 7.8e-6 days/iteration. 2e6 iterations => seems like 15.…

Power draw
262.6 kW
Compute cost
$171,819

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
Open source

training code https://github.com/NVIDIA/Megatron-LM/blob/main/pretrain_bert.py MIT-like license: https://github.com/NVIDIA/Megatron-LM/blob/main/LICENSE

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Frontier model
Yes
Why it is tracked
Highly cited,SOTA improvement

"Our BERT model achieves SOTA results on the RACE dataset"

Record confidence
Confident
Citations
2,766

Sources

Where this record came from and when it was last checked.

Reference
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
Last updated
25 May 2026

What the numbers mean

What this model is

Megatron-BERT was published by NVIDIA, in United States of America, in September 2019. It comes out of industry.

It works in Language, and is recorded as doing language modeling/generation.

This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.

What went into building it

The training run consumed about 2.2 × 10²² FLOP, on NVIDIA Tesla V100S PCIe 32 GB. That figure describes the cost of creating it and has no bearing on how quickly it generates text.

It was trained on about 6,960,000,000 tokens of text.

Its inclusion criterion is highly cited,SOTA improvement.

Answers

Megatron-BERT — common questions

01

When was Megatron-BERT released?

Megatron-BERT was published in September 2019. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

02

What is Megatron-BERT used for?

Megatron-BERT works in Language, and is recorded as handling language modeling/generation. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

03

How much compute was used to train Megatron-BERT?

Around 2.2 × 10²² FLOP, on NVIDIA Tesla V100S PCIe 32 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.

04

What GPU do I need to run Megatron-BERT?

None. Megatron-BERT 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.

05

Is Megatron-BERT open source?

No. Megatron-BERT has not had its weights published, so it exists only as a service controlled by its owner.

06

How many parameters does Megatron-BERT have?

Megatron-BERT has 3.9B parameters. Table 4. 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.

07

Who created Megatron-BERT?

Megatron-BERT was published by NVIDIA, based in United States of America, categorised as industry.

Source

Original publication

Record last updated 25 May 2026

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.