✦ ✦ ✧
Future-Ready · The Stack

GPUs & the AI Boom

The chip invented to draw video-game explosions turned out to be perfect for AI — because both jobs are millions of simple math steps done all at once.
🍳Two Chips, Two Very Different Bets
🧑‍🍳CPU
⚡⚡ ⚡⚡ 4 BIG CORES
CoresA core is one worker inside a chip. A CPU seats just a few experts — each racing through billions of steps a second.
StyleSerial processing — finish one step, then start the next.
AnalogyA few master chefs who can cook anything on the menu — one dish at a time.
💻 Best at: apps, game logic, your whole operating system
🎮GPU
THOUSANDS OF SMALL CORES
CoresA GPU seats a whole stadium crowd — thousands of simpler cores, all working at the same moment.
StyleParallel processing — split the job, solve every piece together.
AnalogyA thousand line cooks, each flipping one pancake at the exact same moment.
🎮 Best at: pixels, matrices, the same math on millions of items
⏱️Serial vs. Parallel: Same Job, Different Time
SERIAL (CPU) — one at a time 1234 5678 🧑‍🍳 done tick 8 PARALLEL (GPU) — all at once 1234 5678 👩‍🍳 All 8 finish in tick 1! Each core is slower on its own — but way more work gets finished per second. That total-work-per-second is called THROUGHPUT ⚡ time
A GPU wins by throughput, not raw speed. But if step 5 needs the answer from step 4, the work can't be split — and that's a job for the CPU.
🔢The Secret: AI Math Is a Grid of Numbers
MATRIX A MATRIX B ANSWER GRID 317 254 609 × 427 193 802 = ? 1 row × 1 column = 1 tiny answer. No cell waits for its neighbors — so ALL of them can be solved at once. ⚡
A photo is really a matrix — one number per pixel. An AI’s “knowledge” is stored in matrices too, some with billions of entries. Training means multiplying those giant grids trillions upon trillions of times — a perfectly parallel job.
🎯 Why nobody planned this
GPUs were built for games — and AI turned out to need the same shape of math.
In 2006, NVIDIA released CUDA, letting scientists program gaming chips to crunch any numbers — chemists, astronomers, and AI researchers pounced.
NVIDIA isn't alone: Google designs AI chips called TPUs, and AMD builds powerful GPUs too.
🕰️How a Gaming Part Took Over the World
1993 NVIDIA founded at a diner 🍳 2006 CUDA lets gaming chips crunch any numbers 2012 AlexNet wins ImageNet on 2 gaming cards 🏆 2023 Global chip shortage — giants wait months in line TODAY First company worth $5 trillion 💵
In 2012 in Toronto, grad student Alex Krizhevsky ran two gaming graphics cards day and night for nearly a week to train AlexNet. It won ImageNet — a contest over a million images — with barely half as many errors as the runner-up. That bedroom experiment lit the fuse on the AI boom.
🏗️From a Bedroom to Warehouse-Scale Brains
HOW MANY GPUs? (bar lengths are squished so the tiny one still shows) Bedroom, 2012 2 gaming cards → AlexNet Frontier, Oak Ridge ~38,000 GPUs supercomputer 🔬 AI cluster 100,000+ GPUs frontier AI 🤖 Memphis site more than 500,000 GPUs 🏢
Frontier AI models train on fleets lashed together with rivers of fiber-optic cable so they act as one giant machine. A single top AI GPU can cost as much as a new car — multiply that by hundreds of thousands and you see why compute is one of Earth's most valuable resources.
⚡ The honest energy story
The cost: a 100,000-GPU cluster can draw over a hundred megawatts — roughly the electricity of a small city — and data centers' share of the power grid climbs every year.
The race back: each new chip generation squeezes far more math out of every watt, and the energy to answer one AI question has dropped dramatically.
🔑Key Terms
🧑‍🍳CPU Central processing unit — a computer's main "brain," with a few powerful cores that race through instructions one after another.
🎮GPU Graphics processing unit — a chip with thousands of small cores, built to draw game images and now powering most AI.
🚶Serial processing Doing tasks one at a time, in order — like a single checkout lane where everyone waits their turn.
⚡Parallel processing Splitting a big job into small pieces and solving them all at once — 30 students each grading one page.
🔢Matrix A grid of numbers in rows and columns — how computers store images and how AI models store what they know.
🏎️Accelerator A specialized chip built to do one job — like AI math — far faster than a general-purpose CPU could.
💪Compute Raw processing power for calculations — horsepower for thinking. Now one of Earth's most valuable resources.
🌍Where You'll Find This in Real Life
🔬Oak Ridge National Laboratory, Tennessee
Frontier, one of the world's fastest supercomputers, uses nearly 38,000 GPUs (made by AMD — NVIDIA isn't the only player) to model cancer treatments, climate change, and new materials. Same parallel-processing idea, scaled to national-lab size.
🎬Game & film studios
The studios behind your favorite games and animated movies run "render farms" packed with GPUs to calculate light, water, fur, and explosions frame by frame — and their job postings now ask for the same GPU skills AI labs hunt for.
🤯 Wildly True Facts
  • NVIDIA was founded in 1993 at a Denny's diner booth in San Jose, California. In 2023 the restaurant installed a plaque marking the spot where a company now worth more than $5 trillion began over cheap coffee.
  • A single top AI chip packs about 336 billion transistors — roughly four times the number of neurons in your entire brain (about 86 billion).
  • In 2023, the cloud company CoreWeave borrowed $2.3 billion using its AI GPUs as collateral — banks accepted stacks of computer chips the way they'd normally accept a house.
⭐
Remember this:
  1. CPUs = a few fast cores for serial work (one step at a time). GPUs = thousands of cores for parallel work (many steps at once).
  2. AI training is mostly multiplying giant matrices — millions of simple, independent calculations — exactly the parallel job GPUs were built to do for video games.
  3. Compute is now one of Earth's most precious resources: frontier AI takes warehouse-scale GPU fleets and serious energy, so engineers race to do more math per watt.
🤔 Think about it
If you had 1,000 helpers who could each do only very simple math, what big problem would you split up — and what could never be split at all?
Frontier AI needs scarce, costly GPUs. Who should get the world's compute — and who decides?
✏️ ClickClass Anchor Chart · GPUs: The Engines Powering the AI Boom
From ClickClass — hundreds of free printables at clickclassedu.com/printables