I recommend watching or listen this video at x1.4 playback speed. If you can't set the speed to this value, then at least watch at x1.25.
Description:
Graphics Cards can run some of the most incredible video games, but how many calculations do they perform every single second? Well, some of the most advanced graphics perform 36 Trillion calculations or more every single second. But how can a single device manage these tens of trillions of calculations? In this video, we explore the architecture inside the 3090 graphics card and the GA102 GPU chip architecture.
Note: We chose to feature the 30 series of GPUs because, to create accurate 3D models, we had to tear down a 3090 GPU rather destructively. We typically select a slightly older model because we're able to find broken components on eBay. If you're wondering, the 4090 can perform 82.58 trillion calculations a second, and then we're sure the 5090 will be even more.
Table of Contents:
00:00 - How many calculations do Graphics Cards Perform?
02:15 - The Difference between GPUs and CPUs?
04:56 - GPU GA102 Architecture
06:59 - GPU GA102 Manufacturing
08:48 - CUDA Core Design
11:09 - Graphics Cards Components
12:04 - Graphics Memory GDDR6X GDDR7
15:11 - All about Micron
16:51 - Single Instruction Multiple Data Architecture
17:49 - Why GPUs run Video Game Graphics, Object Transformations
20:53 - Thread Architecture
23:31 - Help Branch Education Out!
24:29 - Bitcoin Mining
26:50 - Tensor Cores
27:58 - Outro