Syllabus:

1️⃣ Essential AI Knowledge (38%)

AI vs ML vs DL (AMDA)

AI

Machine simulating human intelligence and decision-making

• Example: Alexa, Siri, Self-driving car, Chatbot

Machine Learning (ML)

Ability of machine to learn without explicitly being programmed.

• Example: Spam detection, Fraud detection, Recommendation systems

Deep Learning (DL)

Ability of machines to process data in a way that is inspired by the human brain (by using artificial neural network)

• Focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning.
• Derives rules from the data itself, rather than relying on explicit programming.
• Example: Image recognition, Natural language processing, Speech recognition

Generative AI

Subset of DL that focuses on create new data that is similar to the data that was used to train the model.

• Example: ChatGPT, DALL-E, Stable Diffusion

AgenticAI

Subset of generative AI focused on autonomous decision-making and actions

• Example: Autonomous robots, AI agents, Self-driving cars

PhysicalAI

AI integrated with physical systems (IoT, robotics)

• Example: Smart home devices, Industrial automation, Autonomous drones

Relationship:

 AI
    |-- ML (Supervised, Unsupervised, Reinforcement)
         |-- DL (CNNs, RNNs, Transformers)
             |-- [Generative AI](https://developer.nvidia.com/topics/ai/generative-ai?sortBy=developer_learning_library%2Fsort%2Ffeatured_in.generative_ai%3Adesc%2Ctitle%3Aasc) (LLMs, Diffusion Models)
                    |-- Agentic AI (Autonomous systems, Robotics)
                              |-- Physical AI (IoT, Edge devices)

Term	Key Idea	Exam Trap
AI	Machines simulating human intelligence	Broad umbrella
ML	Learns from data	Doesn’t require manual rules
DL	Neural networks, many layers	Subset of ML
GenAI	Generates new data	Subset of DL

---

GPU vs CPU Architecture

CPU	GPU
Few powerful cores	Thousands of smaller cores
Optimized for sequential tasks	Optimized for parallel workloads
Lower throughput	Massive parallel throughput
Best for control logic	Best for matrix operations

Key Point: GPUs excel at matrix multiplications used in neural networks.

---

Training vs Inference

AI Workflow:

 Data Preperation 
  |--> Model Training 
     |--> Optimization 
         |--> Inference/Deployment

Training	Inference
Model learning	Model usage
High compute + memory	Lower latency focus
Batch workloads	Real-time workloads
Multi-GPU scaling	Edge + cloud deployment

Training = compute intensive
Inference = latency optimized

---

NVIDIA Software Stack (High-Level)

• CUDA → GPU programming platform
• cuDNN → Deep learning primitives
• TensorRT → Inference optimization
• NCCL → Multi-GPU communication
• RAPIDS → GPU data science
• NVIDIA AI Enterprise → Production AI platform

---

Why AI Adoption Accelerated

• GPU performance improvements
• Large datasets
• Cloud scalability
• Transformer architectures
• Pretrained models
• Open-source frameworks

---

2️⃣ AI Infrastructure (40%)

Scaling GPU Infrastructure

Scale-Up

• More GPUs per node
• NVLink
• NVSwitch

Scale-Out

• More nodes
• InfiniBand
• Ethernet
• RDMA

---

Data Center Requirements

Power

• High rack density (30–80kW+ per rack)

Cooling

• Air cooling
• Liquid cooling
• Rear door heat exchangers
• Direct-to-chip cooling

---

Networking Requirements

Important concepts:

• RDMA
• RoCE
• InfiniBand
• East-west traffic
• Spine-leaf architecture

High-speed DC options:

• 100G / 200G / 400G Ethernet
• InfiniBand HDR/NDR

---

DPU (Data Processing Unit)

Purpose:

• Networking offload
• Security isolation
• Storage acceleration
• Free CPU resources

Architecture roles:

• CPU → General compute
• GPU → AI compute
• DPU → Infrastructure acceleration

---

On-Prem vs Cloud

On-Prem	Cloud
CapEx	OpEx
Full control	Elastic scaling
Long-term cost efficiency	Fast deployment
Hardware management required	Managed infrastructure

---

3️⃣ AI Operations (22%)

Monitoring GPUs

Key Metrics:

• GPU utilization
• Memory utilization
• Temperature
• Power usage
• ECC errors
• SM occupancy

Tools:

• NVIDIA DCGM
• Prometheus
• Grafana
• nvidia-smi

---

Cluster Orchestration

• Kubernetes
• Slurm
• Workload scheduling
• Job prioritization
• Multi-tenant isolation

---

Virtualization

• NVIDIA vGPU
• MIG (Multi-Instance GPU)
• GPU partitioning
• Isolation vs performance trade-offs

---