Intelligent computing operations is a full-stack advance from traditional IT ops to LLMOps
A training job worth millions fails on an NCCL timeout; an inference service crashes on out-of-memory; nobody can say how much the GPUs bought at great cost are actually utilized. Who takes responsibility for the engineering problems behind the AI spotlight? Written for AI infrastructure engineers, this book maps the advance from cloud-native to AI-center-native, chaining hardware, platform, optimization, and operations through first principles.
The Full-Stack Operations Path
Key Concepts
Two-tier computing
A collaborative architecture where headquarters runs large-scale training and edge sites serve inference.
It is the mainstream paradigm for state-owned enterprises, deciding the efficiency of model distribution and resource synergy.
MFU
Model FLOPs Utilization — how much of the chips' theoretical peak a training job actually uses.
It is the hardest operational metric, directly answering whether the investment pays off.
AI-native scheduling
GPU/NPU-aware containerization, Device Plugins, Binpack, and related scheduling techniques.
Vanilla Kubernetes can neither see nor manage accelerators; scheduling strategy decides cluster utilization.
Distributed parallelism
Data, tensor, and pipeline parallelism strategies that split large model training.
Thousand-card training inevitably fails; understanding parallelism from an ops view is key to fault localization.
LLMOps
An operations system covering the full model lifecycle: training, inference, monitoring, and billing.
It is the operational leap from selling resources to delivering services, from firefighting to self-healing.
Map of the Book
Part One: Foundation -- Intelligent Computing Hardware and Network Architecture
Part Two: Base Layer -- Building a Cloud-Native AI Platform
Part Three: Core -- Model-Computing Power Matching and Full-Stack Optimization
Part Four: Insight -- Observability and Fault Self-Healing
Part Five: System -- Two-Tier Operations and Future Outlook
After reading, you will understand
- MFU and linearity are the hardest operations metrics
- Vanilla K8s does not fit AI; scheduling must be rebuilt
- Memory and training time are computed, never guessed
- Monitoring spans from the collection layer to the business layer
- Think through billing: per GPU-hour or per token