TPU Deep Dive
CourseReview SlidesExamsSources

Sources

  • SB7All About Transformer InferenceBasics
  • SB7All About Transformer InferenceLinear ops
  • SB7All About Transformer InferenceLatency and throughput bounds
  • SB7All About Transformer InferenceLLaMA-2-13B tables
  • SB7All About Transformer InferenceTricks
  • SB7All About Transformer InferenceSharding inference
  • SB7All About Transformer InferenceEngine design
  • SB7All About Transformer InferenceLatency-bound comms
  • SB7All About Transformer InferenceSpeculative sampling
  • SB7All About Transformer InferenceProblems
  • SB8Serving LLaMA 3-70B on TPUsServing problems
  • P-ESTIPope et al., "Efficiently Scaling Transformer Inference"
  • R-qwixQwix
  • R-tpuinftpu-inference (being renamed vLLM JAX)Qwix config in tpu-inference
  • D-archSystem architecturedtypes per generation
  • D-v5pTPU v5p
  • D-7xTPU7x (Ironwood)
  • G-8ddGupta & Mugazambi, "TPU 8t and TPU 8i technical deep dive"
  • J-hwPallas "TPU Hardware Reference"
Chapter 7. Day 4, first block, 55 minutes

Inference rooflines

ReadQuiz12Flashcards20Teach-back6
Question 1 of 12
Derivation: enter a number
On Ironwood (313 FLOP/B BF16 ridge), weights are stored in fp8 and the math stays in bf16. Above what decode batch are the linear layers compute-bound?
sequences
TPU Deep Dive

A free course on TPU hardware and software for GPU engineers.

Course

Course mapSlidesExamsReview flashcardsMoving off PyTorch/XLA

About

AboutSourcesPrivacy policyTerms of use

Project

Source on GitHubReport an errorYour data

An independent educational project, not affiliated with or endorsed by Google, NVIDIA or AWS. Product names are trademarks of their owners. Sources accessed October 2026; figures marked provisional, derived or unverified are explained where they appear. Written with AI assistance and may contain errors: report one on GitHub.

© 2026 learntpu.dev. Content CC BY 4.0, code MIT.