Full-stack AI systems, taught in Chinese, weighing 10 GB

What it does AISystem is an open-source course — text, video, and slides — that teaches the full-stack anatomy of AI systems. Created by ZOMI from hands-on industry work, it walks through hardware architecture, compiler design, inference engines, and framework internals for senior students, grad researchers, and practicing engineers.

The interesting bit The course explicitly refuses to treat AI chips as isolated “NVIDIA-killers,” insisting instead that usable silicon must co-design with algorithms, compilers, and frameworks. It is also one of the few comprehensive Chinese-language resources covering both international and domestic Chinese AI processors.

Key highlights

• Five modules spanning AI chips (CPU, GPU, NPU, Tensor Core, NVLink), compiler front-ends and back-ends, inference optimization (quantization, pruning, kernel tuning), and framework internals (automatic differentiation, compute graphs).
• Multi-modal delivery: written chapters, Bilibili and YouTube lecture videos, and open PowerPoint slides.
• Hardware coverage includes both global designs (Google, Tesla, NVIDIA) and Chinese domestic processors from Cambricon and Enflame.
• Roughly 17,000 stars and actively maintained, though newer large-model content has been moved to a separate repository.

Caveats

• The repository has ballooned to about 10 GB with raw production assets, so cloning is slow; the maintainers recommend downloading from Releases instead.
• A few compiler sections remain unfinished, and the latest large-model material now lives in a companion repo.

Verdict Worth bookmarking if you are a Chinese-speaking developer or researcher who wants to see the whole stack from silicon to torch.compile. Skip it if you are hunting for a pip-installable library or a quick fine-tuning recipe for LLMs.