Creator explains chips with diagram from @itsmarceliina

The creator explains the fundamental differences between computer chips like CPUs, GPUs, TPUs, and NPUs. She begins by establishing that most people don't understand these differences, then uses a detailed, hand-drawn diagram to illustrate that the core of modern AI is massive matrix multiplication, and each chip is a specialized tool for that math, proceeding to break down the architecture and purpose of CPUs and GPUs.

Creator: @itsmarceliina on Instagram

Video format

Speaker address

Video outline

  1. Pose common knowledge gap
  2. Reveal unifying core concept
  3. Explain familiar baseline
  4. Contrast with specialized solution

Narrative framework

The Scaffolding Explainer

Narrative framework logic

Systematically explaining a category of complex, related items by first establishing a single unifying principle and then using a familiar baseline item as a reference point for comparison.

Topics: Tech, Engineering, Education

Concepts: Breakdown

Formats: Speaker address

Elements: Title Text Hook, Jump Cut

Account types: Personal Brand, Brand

Transcript excerpt

The chips powering AI like GPUs, TPUs, NPUs are some of the most sought after pieces of hardware in the world right no. But most people can't explain the difference between any of them. Let's fix that. What you see here is what every instance of modern AI runs almost entirely on, massive matrix multiplication repeated over and over inside neural networks. Every different chip acronym is just a different specialization of that same math under a different constraint. The chip most people are familiar with is the CPU, the classic general purpose chip. It has a small number of powerful cores built to handle anything, running apps, loading web pages, executing code, responding to a click. It can technically do the matrix math behind AI, but deep learning needs that same operation applied across millions of values simultaneously. And a handful of cores just can't keep up. That's why, for AI, we ended up borrowing a chip from a completely different worse, graphics. GPUs were originally built to render video games via thousands of tiny cores applying the same operation to millions of pixels at once. That structure maps almost perfectly onto neural network math, which needs the same massive

1,715,922 views