Pourquoi l'IA tourne sur des GPU et pas des CPU

Your computer already has an ultra-powerful processor: the CPU. And yet, for AI, it's almost useless. Instead, we plug in a card originally designed for video games. Why? And why is NVIDIA, the manufacturer of these cards, now worth over $3 trillion? In this video, we'll break down why all modern AI runs on GPUs and not CPUs. No math, no technical background required. We'll look at the hardware architecture, Tensor Cores, memory bandwidth, and NVIDIA's real moat—which isn't what you might think. 🔗 Resources: "GPU Puzzles" by Sasha Rush (open source) · NVIDIA technical blog (Hopper/H100 architecture). 🔔 Subscribe to AI Projects for upcoming videos on VRAM, quantization, and LLM inference. Did you think the GPU won simply because it calculates faster? Tell me in the comments what surprised you the most. #GPU #NVIDIA #AI #ArtificialIntelligence #CUDA #TensorCores #DeepLearning #MachineLearning #Hardware #LLM #AIPopularization #AIProjects --- 🔗 SOURCES NVIDIA — Hopper architecture documentation (H100): 4th generation Tensor Cores, ~80 GB HBM3, 3 TB/s memory bandwidth, 1,000 TFLOPs Tensor performance (depending on accuracy). NVIDIA — CUDA history (launched in 2007) and ecosystem (cuDNN, cuBLAS). Krizhevsky, Sutskever, Hinton (2012) — AlexNet trained on 2× GTX 580: GPU switching for deep learning. LLM inference literature: memory-bound nature of the generation (decode) phase vs. compute-bound nature of the prefill phase; role of memory bandwidth. Sasha Rush — "GPU Puzzles" (open-source educational resource on GPU parallel programming).