OPT-175B

OPT-175B is the largest of Meta’s Open Pretrained Transformer models, released in May 2022 and trained at the end of 2021.

It was trained on 992¹ or 1024² NVIDIA A100 80G GPUs running in Azure³ over the course of 56 days.⁴ A 114-page logbook was released by Meta that documents the on-call engineers who kept the training happening during that time, offering a unique view into LLM training at scale.

Efficiency

The model achieved 147 TFLOPS/GPU⁵ and required 4.30E+23 FLOPS total.² Doing some math, this means the total training time was:

$4.30\times 10^{23}\text{ ops}\times \frac{1\text{ GPU}}{147\times 10^{12}\text{ ops}\cdot {\text{ sec}}^{-1}}\times \frac{1}{1024\text{ GPUs}}=2,856,611\text{ s}=793.5\text{ h}=33.1\text{ days}$

The training began on November 5, 2021 and ended on January 6, 2022 for a total of 63 calendar days.⁶ However, Susan Zhang said it was trained over 56 days,⁴ and I haven’t scoured the logbook to see where the missing week went. This means the overall job uptime was between 51.7% and 58.9%.

Footnotes

1. [2205.01068] OPT: Open Pre-trained Transformer Language Models (arxiv.org) ↩

2. metaseq/projects/OPT/chronicles/final_update.md at main · facebookresearch/metaseq (github.com) ↩ ↩²

3. The logbook refers to “CSP” and “cloud,” but the Baselines Logbook refer to the same tools with “cloud” replaced by “azure” (fixmyazure, --full-azure-upload-path). The logbook is also full of Azure-specific terms including blobs. ↩

4. about me | Susan Zhang (suchenzang.github.io) ↩ ↩²

5. Democratizing access to large-scale language models with OPT-175B (meta.com) ↩

6. metaseq/projects/OPT/chronicles at main · facebookresearch/metaseq (github.com) ↩