Phase 1: Digital Foundations¶
Phase 1 · Digital Foundations
Logic, architecture, operating systems, and parallel execution form the base of the whole roadmap.
Learn how computation is represented, executed, scheduled, and accelerated before you try to deploy AI on real hardware.
Layer mapping: Primarily L5 (hardware architecture) and L6 (RTL / logic design), with an important bridge into L3 (runtime behavior) through operating systems and parallel computing.
Role targets: RTL Design Engineer · FPGA Engineer · GPU Runtime Engineer · AI Compiler Engineer · AI Accelerator Architect
Prerequisites: comfort with basic command-line tooling and a working development environment
Start with a small success: Numbers, shapes, and bytes is a CPU-only worked lab with a hand-calculated AI operator, runnable reference, answer guidance, and six correctness checks. It connects the mathematical values to the integer representation and memory budget used throughout this phase.
What comes after: Phase 2 — Embedded Systems, Phase 3 — Artificial Intelligence, then one of the Phase 4 tracks: Xilinx FPGA, NVIDIA Jetson, or ML Compiler
Why This Phase Exists¶
Every later phase assumes you already understand the mechanics of computation:
- how logic turns into hardware behavior
- how processors execute instructions
- how operating systems manage memory and devices
- how parallel programs map work onto CPU and GPU hardware
Use the entry lab and the learning diagnostic to identify which foundations you need. Experienced learners can demonstrate the artifacts and focus on gaps.
Phase Structure¶
| # | Module | What you learn | Why it matters |
|---|---|---|---|
| 1 | Digital Design & HDL | Boolean logic, sequential systems, Verilog, testbenches | The language used to describe hardware |
| 2 | Computer Architecture | ISA, pipelines, caches, memory systems, throughput vs latency | The design logic behind CPUs, GPUs, and NPUs |
| 3 | Operating Systems | processes, memory, scheduling, synchronization, drivers | The software layer that manages hardware resources |
| 4 | C++ and Parallel Computing | SIMD, OpenMP, oneTBB, CUDA, HIP, SYCL | The execution models used by modern AI systems |
Recommended order: 1 → 2 → 3 → 4
If you already know digital logic, you can move faster through Module 1. If you already know OS fundamentals, still do Module 4 carefully; it is the most important bridge into AI hardware work.
What You Should Produce¶
This phase should leave you with visible low-level artifacts, not just notes.
- a small Verilog block plus a testbench
- an architecture explainer or comparison note for CPU vs GPU vs accelerator design
- a debugging write-up around memory, scheduling, or synchronization behavior
- at least one measured parallel program, ideally including a CUDA or GPU profiling artifact
Record those outputs in a simple engineering log, project README, or benchmark note so the work stays visible and reviewable.
Exit Criteria¶
You are ready to move on when you can:
- read basic RTL and explain what hardware it implies
- reason about cache, memory bandwidth, and pipeline bottlenecks
- explain how the OS affects device access and concurrency behavior
- profile a simple parallel workload and describe whether it is compute-bound, memory-bound, or synchronization-bound
That is the minimum base for the rest of the roadmap.
Who Should Prioritize This Phase¶
- Hardware-first learners: do the whole phase in order
- ML engineers moving downward: focus especially on Modules 2 and 4
- Embedded engineers: do Modules 2, 3, and 4 thoroughly even if Module 1 is familiar
Next¶
→ Phase 2 — Embedded Systems · Phase 3 — Artificial Intelligence
Build one contract through the phase¶
Use one small quantized matrix-vector operator as the connecting example. Digital design establishes its representation; HDL gives it a clock/reset and transfer contract; architecture budgets its bytes and cycles; the OS course owns its buffers; parallel computing builds a measured software reference. Keep a separate file for the operator’s shape, signedness, scale, rounding, saturation, and error tolerance.
A phase review should show the same input/output vectors in the mathematical reference, software implementation, and RTL simulation. A fast result with different rounding or an unaccounted copy is a different experiment. Follow the physical AI design studio when assembling the artifacts into a SoC project.
Carry forward: Operator specification, golden vectors, transfer contract, and measured baseline. Continue with the next course and reuse that evidence.