How does a graphics card actually render a 3D scene?

[Odd Arrow]

Rendering starts with raw geometry, a 3D model is genuinely built from a large mesh of connected triangles, each one defined by specific coordinate points floating out in a genuinely three dimensional virtual space. The graphics card first processes all of that raw geometry, transforming and positioning every single triangle correctly relative to the current active virtual camera's specific viewpoint and angle.

Next comes rasterization, where those transformed 3D triangles actually get converted into the specific individual 2D pixels that will genuinely end up displayed directly on your actual screen. The card figures out precisely which specific pixels on that final flat 2D screen each triangle genuinely covers, and in what particular exact order objects should properly and correctly appear relative to each other from the current camera's specific viewpoint.

Then comes shading and lighting, where the card calculates the actual final color for every single individual pixel based on material properties, active light sources, and any relevant applied textures. This step is genuinely where a huge amount of the raw actual computational work happens, since realistic modern lighting effects require running fairly complex mathematical calculations separately for every single pixel on the entire screen.

Modern cards also increasingly use a technique called ray tracing for specific parts of this whole process. Which actually simulates individual real light rays bouncing realistically around a scene the way real actual physical light genuinely does, producing dramatically more accurate reflections and shadows than older simpler simulated lighting approximation techniques ever could, though genuinely at a real significant additional computational cost.

All of this entire complex process has to genuinely happen dozens of times every single second to produce smooth actual motion. Which is precisely why a graphics card contains thousands of small specialized processing cores specifically built to run all of these repetitive calculations simultaneously in parallel, rather than relying on a much smaller number of more powerful but genuinely more general purpose processing cores the way a regular CPU typically does

[Oscar]

TLDR, geometry gets positioned relative to the camera. Rasterization converts triangles into actual screen pixels, shading calculates the final color for each pixel based on lighting, and thousands of parallel cores repeat all of that dozens of times per second

[Linda52]

Makes me appreciate a smooth 60 frames per second game a whole lot more now that entire complex multi step process happening reliably that many times every single second is a real serious and impressive engineering achievement.

Small but real thing

[SilverSurfer51]

The parallel cores versus a smaller number of powerful general purpose CPU cores comparison is quite the cleanest explanation I have personally ever read for why these two specific chip types are actually so fundamentally different from each other.

Held up well

[Kev96]

Really useful writeup!

Always understood a graphics card vaguely handles the actual visuals somehow, but clearly never properly understood the real specific individual steps involved or why the whole process actually needs so many individual small parallel cores

[Courtois75]

My hunch is this whole entire process is exactly why VR gaming really demands such dramatically more raw graphics power. Rendering the whole thing twice simultaneously, once separately for each individual eye, at a high enough frame rate to actually feel smooth and comfortable

[Louise5]

Short version, geometry, rasterization, shading, repeated dozens of times every single second using thousands of small parallel cores rather than a handful of larger more powerful and general purpose ones.

Worth remembering

[Firewall Rabbit]

The triangles as the actual base underlying unit detail quite surprised me a fair bit. Always just assumed 3D models were built from some kind of more particularly complex underlying shape rather than something that basic and simple

[Connor75]

This actually explains why frame rate drops so noticeably hard specifically in lighting heavy scenes.

If shading is really doing that much separate real calculation work for every single individual pixel, that cost would obviously add up incredibly fast. That part surprised me

[LunarDrift Aoife]

Does ray tracing particularly simulate literally every single individual possible light ray in a whole scene. Or does it use some kind of clever shortcut to actually make that calculation realistically practical and fast enough to run? Worth remembering