The 4-Week ML Systems Study Guide 🗓️

<div align="center"> <a href="README.md">🏠 Playbook Home</a> · <a href="cloud/README.md">☁️ Cloud</a> · <a href="edge/README.md">🤖 Edge</a> · <a href="mobile/README.md">📱 Mobile</a> · <a href="tinyml/README.md">🔬 TinyML</a> </div>

With over 800 questions in this playbook, it is easy to get overwhelmed. This 4-week study guide curates a structured path through the core material, taking you from single-node physics up to fleet-scale architectural design.

If you have 30 minutes a day, follow this syllabus.

Week 1: The Physics of Compute & Memory

Goal: Stop guessing about performance and start calculating it. This week builds your "Mechanical Sympathy."

Day 1: The Roofline Model
- Cloud: The Profiling Crisis
- Edge: The NPU Roofline
- Homework: Memorize the Peak FLOPS and Bandwidth for the A100 and H100 from the Numbers page.
Day 2: Precision & Representation
- Cloud: The Underflow Crisis (FP16 vs BF16)
- Mobile: The Accuracy vs Power Trap (INT8 vs FP16)
Day 3: Memory Systems & The KV-Cache
- Cloud: The Sequence Length Trap
- Cloud: The Fragmentation Crisis (PagedAttention)
Day 4: Hardware Architecture
- Cloud: The Sparsity Fallacy
- TinyML: Co-Designing a TinyML Accelerator
Day 5: Pipeline Bottlenecks
- Cloud: The Data Pipeline Stall
- Edge: The ISP/NPU Handoff

Week 2: Scaling Up (Distributed Systems)

Goal: Understand how to split a 70B parameter model across 1,000 GPUs without the network starving the compute.

Day 1: The Memory Wall of Training
- Cloud: The OOM Error (Optimizer State)
- Cloud: The Gradient Checkpointing Boundary
Day 2: Network Topologies
- Cloud: The Cross-Rack Stall (InfiniBand vs Ethernet)
- Cloud: The Oversubscription Choke
Day 3: Collective Communication
- Cloud: The Ring vs Tree Dilemma
Day 4: 3D Parallelism (The Master Class)
- Cloud: The Pipeline Bubble
- Cloud: Dimensioning the 3D Cube
Day 5: Fault Tolerance at Scale
- Cloud: The MTBF Crisis (Young-Daly Equation)
- Cloud: The Straggler Problem

Week 3: Serving at Scale & Edge Deployment

Goal: Move from training to inference. Understand queuing theory, batching, and physical deployment constraints.

Day 1: LLM Serving Metrics
- Cloud: The Serving Inversion (TTFT vs TPOT)
- Cloud: The Black Friday Collapse (Queuing Theory)
Day 2: Advanced Batching
- Cloud: The Serverless Freeze
- Cloud: The Batching Dilemma (Continuous Batching)
Day 3: Disaggregated Serving
- Cloud: The Disaggregated Serving (Prefill vs Decode splitting)
- Cloud: The Decoding Bottleneck (Speculative Decoding)
Day 4: Mobile & App Store Constraints
- Mobile: The App Size Limit
- Mobile: The App Store Model Size Rejection
Day 5: Fleet Updates
- Mobile: The Cross-Version Compatibility Maze
- TinyML: The FOTA Update Risk

Week 4: Ops, Power, and the "Real World"

Goal: Answer the questions that separate standard engineers from technical leads. Security, thermal physics, and silent failures.

Day 1: Thermal & Power Budgets
- Cloud: The Rack Power Budget
- Cloud: The Thermal Throttling Mystery
- Mobile: The Thermal Throttling Cycle
Day 2: Monitoring & Silent Failures
- Cloud: The GPU Utilization Paradox
- Mobile: The Silent Accuracy Degradation
Day 3: On-Device Architectures
- TinyML: Sub-Milliwatt Always-On Wake Word Detection
- Mobile: The 50-Feature Mobile ML Platform
Day 4: Security, Privacy & Silent Failures
- Cloud: The Tokenizer Mismatch
- Mobile: The Federated Keyboard (DP-FedAvg)
Day 5: Visual Architecture Debugging
- Take the Ultimate Test: Try to solve the architectures in Cloud 5. Visual Debugging without looking at the answers.