Manual · Field Guide

The inference engineer's field guide

Where to start, what to learn, and how to verify you actually understand it. A manual for the person who wants to serve models at scale, or just understand one topic deeply.

This is the manual I wish I'd had. It's organized as a roadmap: where to start, what to do next, and how to verify you actually know it. Each topic is a deep dive with its own page, grounded in real numbers and public sources.

The roadmap

Where to start

  1. Learn the fundamentals from the journey posts: attention, KV cache, roofline, kernels. These are the language everything else is spoken in.
  2. Pick a GPU and read its deep dive. Understand what changed from its predecessor, because that's what actually matters.
  3. Understand how GPUs talk: NVLink, NVSwitch, RDMA, GPUDirect Storage. This is the layer that makes or breaks a cluster.
  4. See the cluster at scale: 4000 B200s, disaggregated, with a Mooncake-style cache pool.
  5. Build your own engine in Rust, the way I did. This is the master class.
  6. Look ahead to Vera Rubin, so you're not caught flat-footed.

The topics

GPUs

Playground & workstation

Next gen

Networking

Custom engine

How to use this manual

Each topic page stands alone: read it when you need it. Each is grounded in real numbers and public sources, so you can trace every claim. Start at the roadmap, or jump straight to whatever's biting you.

The rule

If you can't verify it, you don't know it. Every page in this manual is built on numbers and sources you can check.

Back to the blog