【极度硬核】AI时代，英伟达如何成为最大赢家？【CC】 ｜老石谈芯

如果本期视频对你有帮助，就请点赞分享+关注！ 本期视频提到的一些论文在下面： AlexNet： proceedings.neurips.cc/paper/2012/file/c399862d3b9… Transformer (Attention is All You Need) arxiv.org/abs/1706.03762 A3 taejunham.github.io/data/a3_hpca2020.pdf Spatten spatten.mit.edu/ P.S. 关于动态规划，的确说的有问题，也是被NV的翻译误导了，谢谢指出，介绍TSP的时候讲的就是动态规划。

本人長年在AI領域研究，不得不說老石這集內容講得非常的精準到位，就連AI的論文引用都非常的恰到好處，乾貨滿滿

感谢这个视频，目前我在写毕业论文，我用 half-precision 把模型的体积减小，只知道float32变成float16了，又没人告诉我为什么，无意间看了这期视频，解答了我的疑问。

感謝石博士用心整理，希望能看到後續關於CUDA的解說

一年后再来看这个视频，真是好预测！

我做过GPU的AI算子也做过DSA的AI编译器，结论是GPU只是并行度相较cpu高很多有一些向量的指令集，但是相较DSA那种以cube为单位进行计算的专用AI芯片无论性能还是功耗都是完败，所以NV也有DAS的AI加速卡也有做进GPU里边的专门做超分辨率（DLSS）的专用加速单元，结论就是未来是异构的时代GPU做GPU该做的事NPU做NPU该做的事

真的是滿滿的乾貨 !

期待cuda的讲解！我是搞NLP的，平时了解硬件不多，看博主的视频获益匪浅

点赞！蹲一期后续软件和cuda的介绍😆

Like Randy Pausch said in his famous Last Lecture: opportunity + preparation = luck. There are a few salient points worth mentioning: 1. nVidia has been steadily growing and cultivating CUDA, a money pit of a project for the longest time. But it also shows nVidia's belief/vision in GPU for HPC. 2. Alex Krizhevsky and Illya Sutskever were programming wizards who managed to wrangle CUDA into submission so they could run their DL training job on nVidia GPU's, and the rest is history... 3. The coupling of CUDA w/ DL is now de facto. All the major DL frameworks like Tensorflow, PyTorch, etc. all rely on CUDA for GPU acceleration. What about OpenCL, you said? Yeah, you don't hear about it, do you? :) 4. Most of the nVidia's recent commercial success was also due to crytop mining, not from AI/DL A few years ago, a spade of start-ups tried to unseat nVidia in the DL training accelerator market, but can't even find them anymore on Google. Let's see how Jim Keller's Tenstorrent do when it comes out!

超乾的乾貨 讚👍

博主又帅又聪明爱了爱了

謝謝分享專業知識

厉害厉害，专业

谢谢分享

很好，希望下期可以看到cuda，不断学习，努力加油

期待下一期CUDA

超級硬核! super hardcore! 👍

TSP travelling salesperson problem是个NP-complete问题 用dynamic programming可以把O(n!)算法复杂度降低到O(2^n)

能談一下GPU、AI的市場狀況嗎？AMD趕上nVidia了嗎？Intel沒戲了嗎？謝謝您！