Projects

inference-arch-fundamentals

Computer architecture, metal to token.

Working through the hardware that decides how fast a model can be served, from first principles. Each topic gets its own benchmark, measured results and write-up — not a summary of what the textbook says, but what my own machine actually did.

Topics — 3 of 13 complete

T01 Number representation

Quantisation on real model weights — probing weight distributions, finding the outliers that break naive quantisation, and plotting quality against bit width.

T02 CPU pipeline and execution

A C microbenchmark measuring pipeline behaviour — why measured instruction cost diverges from instruction count.

T03 Memory hierarchy

Memory bandwidth, tiling, and a GPU compute/memory crossover experiment — measuring where a workload stops being memory-bound and starts being compute-bound.