How a GPU Really Works: Intro to CUDA, From an iPhone to an H100
This is the full recording of a free live lecture on GPUs and the CUDA programming model, given on 26 September 2026 by Dr. Raj Dandekar (MIT PhD, co-founder of Vizuara AI Labs). It is meant for people meeting GPUs for the first time. ======================================================== Vizuara's GPU trilogy: Part 1, 5D Parallelism: training one model across many GPUs (self-paced): https://5d-parallelism.vizuara.ai/ Part 2, Inference Engineering: serving LLMs fast, from GPU hardware to production (self-paced): https://inference.vizuara.ai/ Part 3, Kernel Engineering: writing the kernels that run on each GPU (live, from 12 October 2026): https://kernelworkshop.vizuara.ai/ You do not need Parts 1 and 2 before Part 3. All three as one bundle: https://vizuara.ai/course-bundle/bundle_10007413 ======================================================== Chapters 0:00 Why kernel engineering, and the story we will follow 3:20 About the instructor 4:15 Why data centres are rising 5:51 Inside a data centre: this is the GPU 7:55 The iPhone Duo and Duo Preview 10:16 What Apple tells you about the iPhone Duo 12:38 Duo Preview: a new picture every 8.3 ms 15:42 The app asks the camera for frames 16:54 Dim light: 8.3 ms to collect light 19:43 Why the phone adds frames: c[i] = a[i] + b[i] 23:15 Where does the maths run? Inside the phone 26:01 Inside the A20 Pro: GPU cores and SRAM 31:00 The same idea at two scales 32:08 Inside an NVIDIA H100 34:08 Threads and blocks, as a school 39:36 The kernel, from the start 42:14 Warps, as rows in a classroom 45:08 Threads, blocks and warps on the H100 51:26 What a kernel engineer writes 52:27 Where the time goes: bytes moved and maths done 55:53 Two kernels, same maths 56:42 Fusing kernels 57:13 What the user feels: stutter or smooth 58:58 From naive to cuBLAS: the chip sets the ceiling 1:00:10 Go deeper: three blogs 1:00:41 The Kernel Engineering Workshop and the GPU trilogy 1:03:40 Questions and answers ======================================================== Every slide and video from the lecture, as a website. You can pause any video, slow it to half speed, or step through the slides with the arrow keys: https://vizuaraai.github.io/inside-the-iphone-duo/ ======================================================== Read further: - The CUDA Programming Model: one image followed through an H100, block by block https://vizuaraai.github.io/cuda-programming-model/ - How a Matmul Kernel Gets Fast: ten kernels, from 1.3% to 93.7% of cuBLAS https://vizuaraai.github.io/cuda-matmul/ - The same climb as an animated book https://vizuaraai.github.io/matmul-book/ ======================================================== #CUDA #GPU #KernelEngineering