Why a Machine Has Two of Each

A computer built for this work carries two processors and two kinds of working memory. That is a strange thing to build. Nobody puts two engines in a car and asks the driver to guess which one to use. So it is worth asking why the duplication exists, because the two answers are not the same answer, and knowing them is the difference between guessing at a spec sheet and reading one.

Start with the processors. The difference between a CPU and a GPU is not that one is newer or better. It is what kind of work each one is shaped for.

A CPU has a small number of lanes, and each one is extremely fast. That design suits work where the second step needs the answer from the first. Opening a file, deciding what to do next, following instructions in order. Most of what a computer does all day looks like that, which is why no machine has ever stopped needing a CPU.

A GPU is built the opposite way. It has thousands of lanes, each slower than a CPU lane, all performing the same operation on different numbers at the same instant. That arrangement is useless for work with a queue in it. It is enormously powerful when nothing has to wait.

Running a model is the second kind. Producing a single word means multiplying an enormous quantity of numbers together, and almost none of those multiplications depends on the result of another. They can all happen at once. Deciding the color of every pixel on a screen has exactly that shape too, which is why hardware designed for video games turned out to be hardware for artificial intelligence. Nobody planned that. The problems happened to match.

Now the memory, where the reason is completely different.

The GPU does not read the file once and remember it. Every single word a model produces requires reading through the entire set of numbers again. That is not occasional access. It is a constant, enormous flow, repeated for every word on the screen.

VRAM sits physically on the card, a short distance from the GPU, connected by something very wide. RAM sits elsewhere in the machine, further away, connected by something much narrower. Both hold things while they are being worked on. Only one of them can deliver at the rate this particular job demands. When the connection cannot keep up, the result is a GPU with enormous computing power sitting idle, waiting for numbers to arrive (APXML, n.d.).

This explains the thing that surprises people who try to run a model on a machine that is slightly too small. It often starts anyway. When a file will not fit in VRAM, some of it can be left in RAM and fetched as needed, an arrangement the card can technically manage (BentoML, 2026). What it produces is not a small penalty. It is somewhere between five and thirty times slower, which in practice means a person asks a question and gives up before the answer finishes (CORSAIR, 2026).

The reverse is also true and is worth knowing. Once a file has been loaded into VRAM and fits there, the narrow connection to the rest of the machine stops mattering almost entirely, because nothing needs to cross it anymore (Glukhov, 2025). The whole job is happening on the card.

So the two answers are these. There are two processors because some work has a queue in it and some does not. There are two memories because the card can only run at full speed on numbers it can reach without leaving home.

Storage sits underneath both and answers a third question, which is where things live when nobody is using them. It is the largest of the three and by far the slowest, and nothing is ever computed there. A file on storage is a file waiting.

Which is why the number that decides everything is VRAM. Not because larger is better in some general sense, but because fitting and not fitting are two different machines. Below the line the work happens on the card at full speed. Above it, the card waits.

A Chromebook has one processor and one memory. Neither of them is the wrong quality. They are the wrong kind, for this one job, and they are perfectly good for every other job the machine was sold to do.

References

APXML. (n.d.). LLM memory bandwidth importance. https://apxml.com/courses/llm-model-sizes-hardware/chapter-3-model-size-hardware-connection/memory-bandwidth

BentoML. (2026, April 15). What is GPU memory and why it matters for LLM inference. https://www.bentoml.com/blog/what-is-gpu-memory-and-why-it-matters-for-llm-inference

CORSAIR. (2026, April 28). Memory for local LLMs: How much RAM do you need? (and when speed matters). https://www.corsair.com/us/en/explorer/diy-builder/how-tos/memory-for-local-llms-how-much-ram-do-you-need-and-when-speed-matters/

Glukhov, R. (2025, June 12). LLM performance and PCIe lanes: Key considerations. Medium. https://medium.com/@rosgluk/llm-performance-and-pcie-lanes-key-considerations-db789241367d

View

The first item is your name. Answer it — that is what puts you on the work now that email collection is off.

Exit ticket · on your own
Could This Class Run Its Own AI? · Day 1 of 7