500-page open-source AI Infra book released: quantitative systems design for the LLM era
udmrzn · x · 2026-09-15
Li Bojie, author of the widely-starred 'Understanding AI Agents' book, has open-sourced his 500-page follow-up 'Understanding AI Infra: Quantitative Analysis and Systems Design' (2.2k stars on GitHub already), complete with a XeLaTeX-typeset PDF, calculation tools, and experiments.
The core thesis: the programming abstraction is moving up from operating systems to model contexts — in the LLM era, models are the OS and AI Infra is the new computer architecture. The book derives system designs quantitatively from hardware constraints and model architectures, citing DeepSeek V4/V4.1's redesigned long-context representation as an example of models adapting to hardware.
Twelve chapters cover model architecture, inference/training workloads, accelerator architecture, operators and runtimes, supernodes, datacenter networking, inference optimization, distributed inference, training systems, resource scheduling, and edge-cloud coordination. Issues and PRs welcome.
More from Infra
- RTX 5090s bought by the pallet at $5,000+ as AI firms outbid gamers — rickasaurus · 2026-09-15
- d-Matrix to plug rival inference chips into Nvidia's NVLink Fusion racks by 2027 — ocean_protocol · 2026-09-15
- Could AI training work like Bitcoin? A compute-voting thought experiment — dbasch · 2026-09-15
- Ornn Launches B300 Payback Forecast as H100 Rental Settles at $2.78/GPU-hour — AnneliesGamble · 2026-09-15
- Claude writes 80% of code at Anthropic, CI jobs up 25x in six months — addyosmani · 2026-09-15
- Jensen Huang: China will have EUV by 2030 and is already there on volume manufacturing — pstAsiatech · 2026-09-15