learntpu.dev hosts TPU Deep Dive, a free, self-paced course on Google's Tensor Processing Units for engineers who already know GPUs. It assumes you know transformers, KV caches, CUDA performance and NCCL. It explains TPU hardware, interconnect, sharding, training and inference performance, and the software stack from the ground up.
How the course is organised
Eleven chapters: five days of two blocks of about 55 minutes each, plus a synthesis chapter comparing TPU, GPU and Trainium architecture. Each chapter has:
- prose with inline citations;
- interactive explainers;
- a 12-question quiz;
- spaced-repetition flashcards;
- a teach-back checklist;
- review slides.
There are day exams, a 30-question final exam, and a slide deck for the whole course.
How it was researched
The spine of the course is How to Scale Your Model by Jacob Austin and colleagues at Google DeepMind (jax-ml.github.io/scaling-book). It is supplemented by Google's TPU papers, the Cloud TPU documentation, JAX, XLA and Pallas documentation, the tpu-inference and TorchTPU projects, and NVIDIA and AWS primary sources.
- Every source is listed on the Sources page with its verification status.
- Every number is recomputed. Each figure in the course is computed by a script in the source repository rather than copied, so it can be re-run when a new chip is announced.
- Uncertain numbers are tagged. "Provisional" marks figures not in an official specification; "derived" marks figures computed from cited inputs that no single source states; "unverified" marks claims no primary source confirms.
- Dates matter. Sources were accessed in October 2026. TPU hardware and software change quickly; TPU 8t and 8i, for example, had been announced but were not generally available at the time of writing.
Use of AI
The course was written with substantial help from AI, specifically Anthropic's Claude. AI was used for research, drafting the prose, writing the quizzes, flashcards and slides, writing the scripts that recompute the numbers, and building the site. The maintainer set the scope, reviewed the material and directed revisions. Every figure is computed by a script rather than taken from the model's memory, and every claim is meant to carry a citation you can check.
There may be errors
Despite those checks, the course may contain errors: a wrong number, a misread source, an outdated API or an oversimplified explanation. AI-assisted writing can state an incorrect thing with confidence, and the field moves faster than any course. Treat the course as a well-sourced study guide, not as an authority, and check figures against the vendor's current documentation before relying on them.
Report an error
If you find an error, a broken link or an outdated figure, please open an issue on GitHub. It helps to include:
- the page address, and the section or question number;
- what is wrong, and what you think is correct;
- a source, if you have one.
Pull requests are welcome too. Corrections are dated in the repository's commit history. The site's code and content are in the tpu-tutor repository, maintained by GitHub user jaganathsr.
License
The course content is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy, adapt and share it, including commercially, as long as you credit "TPU Deep Dive (learntpu.dev)" and link to the license. The site's code is under the MIT License. Quoted third-party material, such as passages from How to Scale Your Model, stays under its authors' terms. Details are in the repository's LICENSE file.
Independence
learntpu.dev is an independent educational project. It is not affiliated with, sponsored by or endorsed by Google, NVIDIA, Amazon Web Services or any other company whose products it discusses. See the Terms of use for trademark notices.