I kept noticing the same thing. Somebody asks a question about their graphics card, gets an answer pitched at their level, and walks away with the symptom explained and the cause still a mystery. The gamer hears "turn textures down". The engineer hears "get a bigger card". The finance person hears "it's the AI boom". All true, all useless for the next question.
The cause is the same in every case: a few gigabytes of memory soldered next to a chip. It is not a hard subject. It just never gets explained once, from the bottom, in a way that lets you stop wherever you need to and come back later.
So I had a go at that.
What it is
One guide, five tiers, and you read as far as you need. The title is Video RAM Killed the Radio Star, because it did, and because the universe is infinite and your framebuffer is not.
It starts with a desk and an artist and ends with a Kubernetes cluster that budgets memory the way a bank budgets cash. The tiers are a spectrum, not a difficulty setting:
- Origin is for anyone who has heard the word and wants it explained with a desk and an artist.
- Vector is for the person who wants to know what is actually on the card and why it is glued down.
- Nexus is for the hobbyist about to spend money, who wants to know why a mini PC can beat a graphics card at one thing and lose at everything else.
- Apex is for the engineer who has to make a model fit, run fast and not fall over, and who needs the arithmetic.
- Zenith is for the architect who has to make all of that work for an organisation: many cards, many teams, a pipeline that runs while everyone is asleep.
Where a section leans on something you might not know yet, there is a short box that explains it and tells you it deserves a study of its own. Those boxes are the map of the guides that come after this one. The KV cache is next.
What is in it
The physics: why a CPU and a GPU cannot share a desk, why bandwidth and not compute sets the speed you feel, and what actually happens when the memory runs out, which turns out to depend on what kind of work it is.
The hardware: GDDR against HBM, why two cards with the same memory can be twice as fast apart, and why a mini PC with unified memory can load a model that the most expensive consumer card cannot and then run it at a fifth of the speed. The 2026 memory shortage gets its own section, because it changed the economics of everything below it.
The arithmetic: why a 70B model needs two H100s, what quantisation actually buys you, and the KV cache calculation that explains why a long conversation eats forty gigabytes before the model itself.
The platform: how Kubernetes learns what a GPU is, what DRA changed in 1.37, how to give a CI job a GPU without it poisoning the next one, how to buy GPUs with code, and how to serve a model without falling over.
Why I bothered
Because the first unexplained word is usually where people stop. Someone hears "memory-bound", cannot picture it, and files the whole subject under things other people understand. I wanted a page where that word gets a box, the box says what it means and that it is worth looking into, and the reader carries on. Whether it works is for you to judge, not me.
Every figure in it is checked against a source dated in the weeks before publication, and the sources are listed at the bottom with links. Some of it will move. Rubin bandwidth, RTX 50 Super timing and when the memory shortage eases are the three most likely to. I have said so where it applies.
It is free, there is nothing to sign up for, and I get nothing out of you clicking it. If it gets one person past the tier they came for, that is enough.
Go and have a look.
