ML Systems
A measured CUDA GEMM comparison, CPU/GPU exercises, and clearly labeled future inference and attention studies.
Choose by what you want to learn
| Project | What exists | First useful task |
|---|---|---|
| CUDA GEMM | Implemented comparison and recorded measurements | Run the checked naive/tiled/cuBLAS comparison or inspect recorded data |
| Kernel Forge | CUDA generation/profiling CLI | Generate a vector-add template and inspect checked timings |
| CUDA memory exercise | Notes and incomplete bandwidth exercise | Complete and check the copy/timing TODOs |
| OpenMP lab | CPU parallelism exercises | Complete a parallel loop and compare correct results across thread counts |
| TensorRT study | Roadmap only | Define export, preprocessing, accuracy, and baseline measurement |
| Triton attention | Roadmap only | Define reference attention and correctness before implementing a kernel |
Learn through a concrete attempt
The tutorial project walks through shapes, commands, correctness checks, and result interpretation. Hardware and shape affect performance; target speedups are not measured results. The complete directory gives each project’s current scope.