Hardware and ML tutorials
Eight consecutive lessons with commands, runnable examples, deliberate failures, expected results, and troubleshooting.
Learn by producing a result
This separate tutorial project connects the tools through small engineering attempts. Each lesson explains the problem, prerequisites, commands, expected output, and limits. Begin with a self-checking hardware example; move to GPU experiments or reference-model behavior when those match your goals.
Guided examples
Who it is for: Learners and engineers who want concrete examples connecting the portfolio tools.
First task: Run the event-counter positive and negative tests, then follow the MCP workflow lesson.
What you can produce: Passing functional checks, an intentional failure, JSON EDA reports, and optional GPU/model exercises.
Current scope: Eight lessons and runnable hardware fixtures. GPU measurements require actual CUDA hardware; model exercises and recorded-data plots have narrower evidence.
Start with the example Source and documentation
Choose a lesson
| Lesson | Practical outcome | Hardware requirements |
|---|---|---|
| 1. First simulation | Generate and verify a counter; deliberately break and restore it | Podman and the Verilog image |
| 2. Sensor-event counter | Check idle, event, reset, and wrap behavior | Podman or Docker |
| 3. MCP workflow | Collect review, simulation, and synthesis JSON reports | Verilog and ASIC images; no LLM account |
| 4. GitHub hardware CI | Catch functional regressions on source changes | GitHub Actions |
| 5. GEMM comparison | Check and compare four FP32 implementations | CUDA; recorded-data plotting needs only Python |
| 6. Kernel generation | Generate templates and interpret measured/model reports | CUDA for execution |
| 7. Spiking-tile reference | Trace threshold, pipeline, and refractory behavior | Python only for the model example |
| 8. Next experiments | Turn an exercise or architecture into a checked first milestone | Depends on the chosen task |
Evidence and contribution
The repository contains runnable positive and negative hardware fixtures and a real MCP SDK client. GPU lessons require an actual GPU for fresh timings. Model traces and recorded-data plots offer accessible routes with narrower claims. Read the validation record before reusing a result, then use the lesson template to contribute another example.