WORLD 01 CUDA ENGINEERING
How a kernel runs.
Ten scenes from one million additions to writing your own kernel: launch, threads, warps, memory geography, tooling and the final challenge.
JOURNEY · TEN SCENES
The full story, in order.
01 · THE PROBLEM
Why a GPU?
One million additions. One core grinds through them — another attacks them all at once.
OPEN →02 · MEET THE GPU
Inside the machine
Six engines called SMs share one die. Click one open and look inside.
OPEN →03 · YOUR FIRST KERNEL
What happens when you launch?
The full machine, live. Edit the launch config, scrub time in either direction.
OPEN →04 · THREADS
A thread is one copy of your code
Slide from one thread to a thousand and watch the workload change shape.
OPEN →05 · THREAD INDEXING
Where does this thread work?
Every thread computes its own global index — the equation and the proof, together.
OPEN →06 · WARPS & DIVERGENCE
One warp, two paths
A branch splits 32 lanes into two masks — executed one after another, then merged.
OPEN →07 · MEMORY GEOGRAPHY
Memory is a place, not a bucket
Predict the winner, then watch coalesced versus strided access fight it out.
OPEN →08 · THE DESCENT
Peel back the abstraction
Program → kernel → grid → block → warp → thread → SM → memory.
OPEN →09 · DEBUG IT
Break it, see it
A broken kernel writes to the wrong places — the visualization is your debugger.
OPEN →10 · FINAL CHALLENGE
Write the kernel
Map every thread to its own element. Run. Tests decide.
OPEN →