Series 1 · Episode 2 · 19 Oct 2026
AI on Silicon
LLM inference, measured on real GPUs.
I take a real model, a real inference engine and a real GPU, and explain what the hardware is actually doing, one idea per episode. Every number is my own measurement, every claim names its source, and every episode comes with a Jupyter notebook that reproduces it.
1
Measured
Real models on real inference engines, profiled on real GPUs with tools like Nsight Systems and Nsight Compute. No borrowed benchmarks.
2
Sourced
Each episode ends with where every number comes from: the run, the file, the formula, the datasheet.
3
Reproducible
Every episode links a Jupyter notebook that prints its numbers on any laptop or in Colab, or re-measures them on your GPU.
Series 1 · 2 episodes
LLM inference through the lens of a GPU
Series 1 · Episode 1 · 18 Oct 2026
To an H100, Qwen3.8-27B isn't a model. It's a loop of 902 kernels.
Get the next episode
One email when a new episode is out, plus one figure that doesn’t fit on LinkedIn. No spam.