Google DeepMind Releases 'How To Scale Your Model': A Systems View of LLMs
philhchen · x · 2026-08-18
Google DeepMind has released a new book, 'How To Scale Your Model', providing a systems perspective on LLMs running on TPUs and GPUs. It covers Roofline analysis, inter-device communication, parallelization strategies for training and inference, and cost estimation. The book aims to demystify the science of scaling models for developers optimizing performance at massive scale.
More from Infra
- OCP's Silicon Photonics Vision Roasted as Per-Token Energy Nears 1 Joule — jwt0625 · 2026-08-18
- OpenRouter token usage exploded from 3.73T to over 75T in one year — scaling01 · 2026-08-18
- Crusoe in IPO Talks with Banks, Valued at $35B in Fundraise — nwilliams030 · 2026-08-18
- MiniMax H3 Dynamic Workflow: Running on 16GB VRAM with 4-sec Steps — DaExChef · 2026-08-18
- MiniMax H3 Quality Degrades After ComfyUI Update — ASK_ABT_MY_USERNAME · 2026-08-18
- Dev Plans Free Community Access to Local Qwen 3.8 27B at 56 Tokens/sec — No_Run8812 · 2026-08-18