Phase 5: Advanced Topics and Specialization¶
Phase 5 · Specialization
Turn broad roadmap coverage into a role-shaped body of work with a flagship artifact.
Go deep in one part of the stack until your work looks differentiated, not just broad.
Layer mapping: Depends on the track. This phase spans L1 through L8, but each specialization should stay coherent around one professional niche.
Role targets: GPU Infrastructure Engineer · HPC Engineer · ML Systems Engineer · Robotics Engineer · Autonomous Vehicles Engineer · AI Accelerator Architect · Advanced Edge AI Engineer
Prerequisites: one primary Phase 4 track completed to a project-capable level: Xilinx FPGA, NVIDIA Jetson, or ML Compiler, plus whichever earlier phases your track depends on most heavily.
What comes after: combine specialization evidence in a model-to-machine capstone. For the physical-AI goal, the endpoint is a versioned VLA workload running through a compiler/runtime and hardware prototype, with Linux, sensor, wireless, and control interfaces documented.
Why This Phase Exists¶
Breadth gets you into the stack. Specialization is what makes you recognizable inside it.
This phase exists so you can choose one area and go from:
- “I understand the system”
to
- “I can do difficult, valuable work in this niche”
That is the difference between a learner and a differentiated candidate.
Phase Structure¶
| Track | Focus | Best for | Guide |
|---|---|---|---|
| Track A — GPU Infrastructure | multi-GPU systems, interconnects, distributed runtime, vendor ecosystems | cluster and infrastructure work | Guide |
| Track B — High Performance Computing | CUDA-X libraries, performance engineering, large-scale acceleration | systems and performance roles | Guide |
| Track C — Edge AI | efficient deployment, real-time constraints, edge optimization | edge deployment specialists | Guide |
| Track D — Robotics | ROS 2, autonomy software, robot systems integration | robotics and embodied AI | Guide |
| Track E — Autonomous Vehicles | perception, safety, deployment, debugging real vehicle stacks | AV systems and applied autonomy | Guide |
| Track F — AI Chip Design | accelerator architecture, dataflow, custom compute design | architecture and chip-design paths | Guide |
| Track G — ML Systems Engineering | training systems, inference runtimes, GPU scheduling, distributed serving, compiler/runtime work | AI infrastructure and MLSys roles | Guide |
Recommended rule: choose Track F as the integration destination for the custom-chip goal, then draw evidence from Robotics/VLA (Track D), deployment (Tracks C/G), and performance/infrastructure (Tracks A/B) as the workload requires. You do not need to finish a cluster-training track before building a small inference accelerator.
What You Should Produce¶
Phase 5 should end with a project that is clearly role-shaped:
- a multi-GPU or performance-engineering runbook
- an MLSys benchmark suite or runtime prototype
- an edge deployment case study with hard constraints and metrics
- a robotics or AV system analysis tied to real software stacks
- an accelerator architecture note, prototype, or design-space study
This phase should not end as a pile of bookmarks. It should end as evidence of specialization.
Exit Criteria¶
You are using this phase correctly when you can:
- explain why your chosen specialization matters in the AI stack
- show artifacts that are harder and narrower than your Phase 4 outputs
- connect your specialization to real target roles and actual responsibilities
- point to one body of work that represents your “signature direction”
How To Choose A Track¶
- Choose the track whose daily work you actually want, not the one that merely sounds prestigious.
- Prefer the track that compounds what you already built in Phase 4.
- Avoid trying to complete every specialization; depth matters more than total coverage here.
Next¶
Choose one specialization guide above and turn it into a flagship project.
Study method: use the shared lesson cycle and completion rubric with this course's existing labs. Predict first, test one variation, retain a failure case, and revisit the self-check after two days and one week.
The physical-AI path through specialization¶
| Step | Input from earlier study | Work here | Artifact used next |
|---|---|---|---|
| Define useful behavior | Sensors, ML, embedded systems | Track D VLA and simulation | Timestamped observation/action contract and task-success baseline |
| Locate the cost | Phase 4 deployment/compiler skills | Tracks C/G, selected A/B experiments | Operator shapes, precision requirements, memory/latency trace |
| Build one hardware slice | Digital design and compiler fundamentals | Track F simulator and RTL/FPGA | Verified command ABI, driver path, resource/latency evidence |
| Integrate the machine | Embedded Linux, wireless, control | Track F with D/E system analysis | Closed-loop result, stale-input/link-loss response, reproducible image |
Return to a prior step when the evidence changes the design. If a model's action head dominates latency, optimizing an unrelated text-decoding kernel does not advance the machine. If CPU fallback dominates after acceleration, the next lesson is compiler coverage or data movement. Record the maturity achieved: simulation, FPGA prototype, physical-design estimate, and measured silicon are different outcomes.