From Portable AI Lab to Frontier Research
summary
Builds a complete, inspectable research workflow on a MacBook + tinygrad and follows a tensor op from Python through compiler optimization to a generated Metal kernel, measures per-kernel times and GFLOPS (single test shows ~2000 GFLOPS and a tiny JIT 4x speedup), inspects scheduling decisions and uses a beam-style automatic search to find faster kernels, runs the same stack on an external GPU over USB4 dock, and demonstrates that laptop-first GPU exploration and reproducible kernel hacking actually works.
AI-written summary · about
chapters (11)
- 0:00Introduction and background
- 1:10AI research landscape
- 2:55GPU bottlenecks and limitations
- 8:43From model to GPU execution
- 11:34Researching on a laptop
- 14:50External GPU with TinyGrad
- 20:32Lazy graph and kernel optimization
- 23:58TinyGrad philosophy and simplicity
- 29:00Portable AI research with external GPUs
- 31:09Research cycle and next steps
- 31:52Q&A session
