Forget the brand names for a second, here is the mechanism, and conveniently AMD’s FY2021 10-K lays out the vocabulary. The filing describes customers using “an APU, GPU, SoC or a combination of a discrete GPU with one of the other foregoing products working in tandem.” Three acronyms, one useful map.
The way this actually works: a GPU is a graphics processing unit, but the word “graphics” is historical baggage. What it really is, is a chip with thousands of small cores that do the same arithmetic on many pieces of data at once. Neural networks are, under the hood, enormous piles of that same parallel arithmetic, which is why the gaming chip became the AI chip.
An APU (accelerated processing unit) fuses a CPU and a GPU on one die; an SoC (system on a chip) packs many functions together. A “discrete GPU” is a standalone GPU you add for extra parallel horsepower. The 10-K’s phrase “working in tandem” is the key idea: AI systems mix a general-purpose CPU brain with a parallel GPU workhorse.
Why read this in a 10-K rather than a textbook? Because the filing is the company’s own, legally careful definition of what it sells, with no marketing incentive to overstate. When AMD tells the SEC its customers combine these parts, it is describing the actual architecture of the machines that train and run AI models, stripped of hype.
As a plain-language marker for early 2022: the chips powering AI are not magic, they are CPUs and GPUs combined in a few standard ways, and a company’s own annual report is one of the cleanest places to learn the difference. Keep this vocabulary; every AI-hardware story reuses it. Filing data and the evidence index via EdgarBeast.