Make AI run fast and cheap on real hardware
- The experiment
- Evaluate how compressing a model (quantization) changes speed and quality on my Mac, then compare with my NVIDIA machine.
- Builds toward
- model size and memory, KV cache, batching.
Research
Four tracks. One at a time.
Tracks are done one at a time. Results are published only when I can reproduce them.
GPU programming from scratch
- The experiment
- Write a matrix multiplication, first naive, then optimized, and measure each speed-up.
- Builds toward
- how a GPU really works, and why track A behaves the way it does.
AI on small devices
- The experiment
- Run a small model on a Jetson and measure speed and power use.
- Builds toward
- the limits of memory, speed and power on the edge.
Parallel computing for science
- The experiment
- A toy simulation (particles or molecules) sped up on a GPU. It is a toy, not real drug discovery — a first step toward that world.
- Builds toward
- thinking in parallel, speed-up and its limits.