nvidia-NVIDIA Keynote at COMPUTEX 2023

1. nvidia-NVIDIA Keynote at COMPUTEX 2023-Huang Renxun’s speech on new concepts

Video material: NVIDIA Keynote at COMPUTEX 2023

• 1. Convai avatar framework and virtual avatar tool maker 00:10:50-00:12:00around
• 2. Product iterators, first Kepler, then Maxwell, Pascal, Volta, Ampere, and Hopper
• 3. Short video of making H10000:31:05-00:32:05
• 4. nvlink、InfiniBand、Quantum
• 5. Chip-to-chip serial-parallel transceivers (SerDes)
• 6. Distributed computer operating system Magnum I/O
• 7. New situation in the computer industry: Software programming is no longer done by computer engineers, but by computer engineers with the assistance of AI supercomputers. These AI supercomputers are a new type of factory, namelyAI工厂
• 8. In 10 years, we will increase the throughput scale and the total throughput of the entire stack by 1 million times, 1,000 times in 5 years, and 1 million times in 10 years. Compared with Moore's Law, this speed requires twice as many doubling cycles. The question now is what can you do if your computer is 1 million times faster?
• 9. We can now guide AI to generate various information text-to_text, text_to_image. The most important thing is that now we can realize different forms of information conversion 文本_to_蛋白质, 文本_to_化学物质, image_to_3d, image_to_2d, image_to_文本字幕, video_to_video, which is the first in the history of software technology.
• 10. Let's tell AI, this is an attempt to convert Google text to music text_to_music: I want traditional Taiwanese style music to be calm and peaceful, like deep in the forest at dawn, bathed in drizzle on a warm day. in 00:45:14-00:;
• 11. I'm at Computex Taipei and I'll make you like me the most, sing along with me, I really like NVIDIA00:46:01-00:46:45
• 12. Digital biology is about to usher in its revolution, and before that we had researchers Synopsizeand Cadenceresearchers to help us build the tools so that we can produce great chips and systems. Now for the first time ever we're going to have computer-aided drug discovery tools that will be able to manipulate and manipulate proteins and chemicals and understand disease targets and try all kinds of chemicals that weren't even imagined before, and this is a huge field.
• 13. adobeIt is being Fireflyintegrated into their creative applications to provide professional ethics training and is also convenient for artists to use; the Tabnineknowledge base accumulated by millions of developers is being used to accelerate application development, speed up debugging, and promote popularization; if a Architects who want to redecorate their homes Planner 5Dcan instantly turn 2D floor plans into 3D models; AnHorn Medicinesdifficult-to-treat diseases are being conquered in Taiwan
• 14. Grace HopperNow in full production, DGX GH200it's a giant GPU
• 15. Nvidia's future work direction; one: we already have it Grace Hopper; two: we use NVLink8 Grace Hopperconnected, which is NVLink Switch, 8 Hopperconnected into 3 switch trays to form 8 Grace Hopperdelivery nodes, each of Grace Hopperwhich has 900GB/sa base width of 8 connected to each other. , to achieve ; three: the computing power obtained 600GB/sby connecting 32 such structures (connecting 256个 Grace Hoppersuper chips) through a switch layer ( the computing power: many countries and regions are working hard on research ), this goal has been achieved recently . The engine is formed and each GPU gets memory, . In the video1 ExaFLOPS1 ExaFLOPSE级计算1 ExaFLOPS Transformer144TB这可不是144TB分布式内存，而是144TB的互联内存1:00:28-1:00:50
• 16. In 1964 IBM 发布了 System 360, AT&T展示了全球第一部可视电话encoding, compression, streaming and decoding were implemented through copper telephone lines and twisted pairs. To this day, this process has generally not changed; future communication processes will no longer be compressed or streamed. , decompression, but instead becomes perceptual streaming and reconstruction, which can be generated using a variety of different means. Can generate 3D forms, can regenerate your language into another language, now we have a translator that can take us around the world
• 17. NVIDIA Maxine 3D,virtual image,
• 18. Generative AI, cloud, video and images, EDA, SDA
• 19. InfiniBandBased on RDMAtechnology, this is a flow control technology, which is a lossless network method; Spectrum-4switches 128个400Gb/s端口，51.2Tb/s的交换能力have1000亿个晶体管，尺寸为90mm*90mm,底部有800个焊球，这是一个功耗500瓦的芯片，这款交换机的功耗是2800瓦，48块PCB将交换机连接在一起
• 20. BlueField-3是400Gb/s的网卡Directly connected to Spectrum-4the switch, connected network cables, and running some software, these four constituteSpectrum-X
• 21. NVIDIA AI Foundationsservice, which is a cloud service that provides access to NVIDIA's AI expertise, allowing you to train an AI model, and 我们提供有监督微调SFT、信息护栏、专有知识库和强化学习RLHF技术we can deploy this model to run NVIDIA AI Enterpriseon
• 22. The next stage of AI: the combination of AI and digital twins,Omniverse
• 23. Adobe FireFly、Getty、Shutterstock
• 24. Used toomniverse train ai, and then use the ai edge deployment system Metropolisto deploy ai
• 25. WPPProducing 25% of the world's advertising, 60% of the world's largest companies are already their clients, they have created a video showing how they use this technology video1:40:57-1:42:20
• 26. How electronics manufacturer Pegatron is using NVIDIA AI 和Omniversefactory digitization to generate Omniverse Reolicatordefects PCBA合成数据集, NVIDIA预训练模型NVIDIA Tau, and real-time inference—NVIDIA DeepStream1:43:20-1:45:40
• 27. The Robotics Platform NVIDIA Isaac AMRis now available as a reference design for anyone who wants to build a robot, Robotics Reference Stack Nova Orin—1:48:00-1:49:38