Producing Intelligence at a Glance: How Llama 3 405B Was Made, End to End
One AI factory, one model, nine stages. This page puts the whole Producing Intelligence series on one map, following Meta's Llama 3 405B from the software that runs a 16,384-GPU cluster to the operations team that keeps it alive. Click any stage for its key number, the decision Meta made, the main risk and the tools involved, then jump to the full article.
Explore the pipeline ↓ Start at Part 1 →🏭 The pipeline
Read from the top down: infrastructure and data come first, the training run turns compute into a base model, post-training and evaluation turn it into a product, and serving and operations keep it running.
🗓️ The Llama 3 405B timeline
The order of work, from cluster to release. Meta published dates only for some milestones; the rest are shown in sequence without dates.
⚡ Where the compute went
The published GPU-hour figures, on the same scale. Pre-training dominates the published numbers; Meta did not publish GPU-hours for post-training, evaluation or serving, so those appear as labelled gaps.
Sources: Llama 3 paper and model card (30.84M H100-hours for 405B pre-training; 39.3M for the Llama 3.1 family; 8,930 t CO₂ location-based). See Part 3.
🧭 Decision map
At every stage Meta chose one option among several. The table shows the choice, the main alternative and why it fitted Llama 3. Click a row to open the article.
⏱️ Follow one GPU-hour
What a single H100-hour in Meta's cluster went through during the 405B run, from the power socket to a token in a user's chat.
🔢 Key numbers
One headline figure per part. Each tile links to its article.
🧭 Reading tracks
Every part stands alone. If you do not want to read all nine in order, pick a track.
📖 Glossary
The layer below this series is the hardware, energy and money that make the factory possible: 🖥️ see the AI Compute Stack at a glance →