Go Summarize

輝達GTC大會黃仁勳演講 秀最強AI晶片架構Blackwell|完整版中.英CC字幕|TVBS新聞 @TVBSNEWS01

TVBS NEWS2024-03-19
55#55台#55頻道#tvbs#tvbs新聞#新聞直播#新聞#直播#NEWS#TaiwanNews#NewsLive#24H直播#即時新聞#十點不一樣#氣象#天氣#24小時直播#專題新聞#TVBS#國際#娛樂#政治#56台#NEWS LIVE#最新#Taiwan News#Taiwan#即時#烏俄#中國大陸#烏俄軍#烏克蘭總統#烏克蘭#台股#基輔#聯合國#北約#澤倫斯基#普欽#台積電#疫情#空軍#裴洛西#中美台#解放軍#中國#美國#台灣#眾議院#US#China#軍演#習家軍#蔣萬安#柯文哲#蔡英文#朱立倫#蘇貞昌#郭台銘#布林肯#侯友宜#賴清德#韓國瑜#性騷#以巴#以色列#蕭美琴#不分區立委#Taiwan news#李善均#趙少康#立法委員#黃國昌#王世堅#謝龍介#黃珊珊#晚安小雞#瘦肉精#年代旅遊#富國島#辣椒粉#中國兩會#李強
17K views|3 months ago
💫 Short Summary

Nvidia discusses the transformative impact of accelerated computing and AI in various industries, showcasing advancements in chip architecture, GPU communication, and generative AI models like Blackwell and Cordif. Partnerships with industry leaders drive innovation, with a focus on AI-enabled digitalization in manufacturing and healthcare. The future of software building involves AI systems like NIMs for enhanced productivity. Robotics advancements include AI-powered systems for adaptive routes and human-like training data. Nvidia's commitment to developing specialized infrastructure and software applications highlights the ongoing evolution towards a new industrial revolution centered on generative AI and personal intelligence.

✨ Highlights
📊 Transcript
Nvidia and Blackwell: Transforming Industries with Accelerated Computing.
10:56
Blackwell is a platform that illuminates galaxies, while Nvidia creates GPUs.
The Hopper platform has revolutionized the world, and Nvidia AI is making an impact in healthcare and transportation.
Diverse researchers and industries are utilizing AI for innovation, showcasing the significant impact of accelerated computing.
The computer industry is driving fundamental transformations across all sectors, representing $100 trillion of global industries.
Nvidia's journey from 1993 to the emergence of generative AI in 2023, leading to the creation of a new industry.
11:50
Development of revolutionary computing models like Cuda and DGX One, impacting AI capabilities.
Shift towards producing software using AI factories and generating tokens at a large scale.
Comparison to the industrial revolution's impact on electricity generation.
Highlight on Nvidia's focus at the intersection of computer graphics, physics, and AI in virtual world simulation.
Nvidia emphasizes the importance of accelerated computing for sustainable scaling and cost reduction in various industries.
19:10
Nvidia aims to increase computing scale for digital twins in product simulation.
Nvidia partners with companies like Ansys, Synopsis, and Cadence to accelerate ecosystems.
Nvidia accelerates computational lithography with domain-specific libraries and applies generative AI to semiconductor manufacturing.
Collaboration with Cadence involves building a supercomputer for fluid dynamic simulation at a larger scale.
Advances in supercomputing technology and the integration of AI copilots in design processes.
23:01
Cadence Millennium's digital twin platform connected to Omniverse to accelerate CAE, EDA, and SDA industries.
Exponential growth of large language models with significant computational requirements for training.
Emphasis on the need for bigger GPUs and the development of supercomputers like EOS in 2023.
Importance of building technology for progress in the world.
27:15
Emphasis on developing chips, systems, networking, and software for advancement.
Acknowledgment of complexities in creating systems with efficient energy use and computational power.
Mention of advancements in AI models like chat GPT and need for larger models trained with multimodal data.
Significance of using synthetic data generation and reinforcement learning to enhance AI capabilities.
Introduction of Blackwell, a revolutionary GPU system named after mathematician David Blackwell.
The Blackwell chip boasts 208 billion transistors and offers ten terabytes per second, eliminating memory locality and cache issues.
38:19
The chip is designed for easy integration into existing infrastructure, providing efficiency and cost-effectiveness.
The Blackwell system includes a Grace CPU with super fast chip-to-chip links for memory coherence and seamless collaboration.
Additional features like a transformer engine and dynamic rescaling aim to push the limits of physics and enhance performance.
Importance of precision in numerical formats and the role of artificial intelligence in probability.
39:21
Design of a smaller ALU and the necessity for mathematics to maintain precision and range.
Challenges of running computations on thousands of GPUs, including synchronization and information sharing.
Advancements in transformer engines, NV link technology, and the utilization of all reduce and all gather methods for GPU collaboration.
Emphasis on encryption for data security and implementation of compression engines for efficient data movement.
New computers designed for efficiency and performance in training and inference tasks.
43:37
Emphasis on shift from retrieval to generative AI for energy, bandwidth, and time savings.
Future of computing centered on generative AI with specialized processors.
Notable exponential increase in computational power and need for larger GPUs to meet demands.
The NV link switch is a revolutionary chip with 50 billion transistors and four MV links.
49:17
Each MV link has a speed of 1.8 terabytes per second, allowing for full-speed communication between GPUs.
The chip drives copper directly, resulting in cost-effective systems and the ability to create massive GPU clusters.
The DGX system boasts 720 petaflops of power, equivalent to one exaflops for training purposes.
It features a 130 terabytes per second MV link spine and liquid cooling to manage the 120 kW power consumption, achieving high bandwidth without expensive transceivers and saving 20 kW for computation.
Evolution of GPUs has led to complex models with 600,000 parts and weighing 3000 pounds.
54:10
Training large AI models like GPT requires 8000 to 2000 GPUs and 15 MW of energy over 90 days.
The focus is on reducing costs and energy consumption while enhancing computational capabilities for AI training.
Inference tasks for large language models, like chatbots, require supercomputers to handle real-time interactive applications with trillions of parameters.
Importance of distributing work across multiple GPUs for high throughput and interactivity in generating tokens.
57:55
Balancing throughput and quality of service is crucial when optimizing performance across different GPU configurations.
Nvidia's GPUs, programmable through CUDA, allow for exploration of distribution strategies to maximize performance.
Comparison of inference capabilities between Blackwell and Hopper for generative AI models, with Blackwell outperforming by a factor of 30 for large language models like GPT.
Potential performance improvements in chip architecture and GPU communication.
01:02:25
Introduction of the new transformer engine and MV linked switch for faster data sharing, beneficial for AI applications.
Blackwell system highlighted for generative AI and revenue generation in data centers.
Anticipation for Blackwell's launch emphasized, with partnerships with AWS and Google for GPU and AI integration.
Development of a secure AI GPU and collaboration with Amazon Robotics and AWS Health for accelerated computing solutions.
Collaboration between GCP, Oracle, Microsoft, and Nvidia is accelerating data processing, robotics, and AI services.
01:07:56
Digital twins are becoming popular for optimizing processes and reducing costs through virtual replicas of physical environments.
Wistron used Nvidia's Omniverse SDKs to create digital twins of factories, enhancing worker efficiency and online deployment speed.
Nvidia and its partners are spearheading AI-enabled digitalization in manufacturing, ushering in a new era of technology integration.
AI technology breakthrough revolutionizes software development.
01:11:32
Data compression allows for recognition and understanding of text, images, and sounds.
Complex structures like proteins, genes, and brainwaves can be digitized and analyzed for pattern recognition.
Generative AI has potential applications in climate prediction and creating a digital twin of Earth for forecasting.
Nvidia's Cordif AI model improves storm tracking resolution from 25km to 2km to enhance speed and energy efficiency.
01:15:04
Cordif combined with forecastnet enables accurate regional weather forecasting, reducing storm impacts.
Nvidia plans to make Cordif available worldwide through the Earth 2 inference service.
Nvidia also applies AI models to analyze genetic and protein sequences in healthcare advancements.
Models like AlphaFold are revolutionizing protein reconstruction to aid researchers in virtual screening for new medicines.
Utilizing NIMs for protein structure prediction, molecule generation, and docking enables quick candidate molecule generation and screening.
01:19:21
Molmim can optimize for desired properties, including binding to target proteins, through custom applications.
The approach results in high-quality drug-like molecules, increasing the likelihood of developing successful medicines faster.
Bionemo is revolutionizing drug discovery by offering on-demand microservices for powerful workflows like de novo protein design and guided molecule generation.
The future of software building involves the use of AI systems called NIMs.
01:23:13
NIMs can handle tasks such as understanding SAP, retrieving information, performing calculations, and numerical analysis.
Companies like Nvidia have implemented NIMs to enhance productivity, like creating chatbots and aiding chip designers.
NIMs can be customized and programmed to perform specific tasks, improving efficiency and output within an organization.
Nemo microservice aids in data curation and preparation for AI training, evaluation, and deployment.
01:27:39
Three key pillars of Nemo's work involve inventing AI technology, developing tools for modification, and offering infrastructure for fine-tuning and deployment.
Nemo strives to become an AI foundry akin to TSMC for chip building.
Nemo can comprehend proprietary information and construct a vector database for both structured and unstructured data.
The database enables efficient communication with encoded data like PDFs, enhancing data retrieval and interaction for software teams.
Collaboration between Nvidia AI Foundry and major companies to build AI assistants and copilots.
01:31:26
Partnerships with companies like SAP, Cohesity, Snowflake, and NetApp to enhance customer service operations and data management.
Dell's role in constructing AI factories for enterprise-scale systems, signaling the importance of investing in AI infrastructure.
Advancements in AI and robotics through the use of three computers: AI computer, video-watching AI, and simulation engine for physical learning.
01:35:36
Systems such as DGX, AGX, and Jetson are utilized in this next wave of AI and robotics.
Reinforcement learning human feedback enables robots to align and adapt to physics laws with physical feedback.
The simulation engine called Omniverse in the Azure cloud offers a virtual world for robots to learn.
These systems rely on algorithms and collaborate to enhance AI and robotics capabilities.
Robotics building acts as an air traffic controller for autonomous systems.
01:37:52
Warehouse provides real-time redirection for robots and people.
Digital twin of the warehouse used for simulation with AI agents and AMRs.
Operators can interact using natural language and receive immediate insights.
Future envisions software-defined facilities with continuous improvement of digital twins and AI models.
Nvidia Omniverse integrating AI with chat USD for communication.
01:41:46
Users can interact in English and receive USD responses, enabling semantic searches and collaboration in 3D design and AI generation.
Siemens partnering with Nvidia to connect their industrial engineering platform to Omniverse, enhancing data interoperability and physics-based rendering for large-scale projects.
Collaboration aims to create an intuitive digital twin that reduces errors and costs, increasing productivity in design, engineering, and manufacturing processes.
Nissan integrates omniverse into their workflow for seamless collaboration.
01:46:11
The integration enables virtual exploration and design creation, with a focus on robotics and autonomous systems in the automotive industry.
NVIDIA unveils Thor, an AV computer adopted by BYD, and Jetson, a robotics computer for developers.
The company prioritizes CUDA compatibility and ecosystem support for developers, unveiling the advanced Isaac Perceptor SDK for autonomous and programmable robotics.
The future of robotics involves perception and programming waypoints for adaptive routes.
01:53:23
The Isaac Perceptor offers advanced vision, odometry, 3d reconstruction, and depth perception.
CUDA accelerated motion planning allows for dynamic obstacle avoidance.
The next step is humanoid robotics, leveraging human-like training data for enhanced productivity in various industries.
The development of humanoid robots like Thor, equipped with transformer engines, showcases ongoing advancements in robotics technology and potential for increased efficiency in work environments.
Nvidia's project group is working on robot learning with a foundation model and Isaac Lab for training in simulation.
01:57:10
OSMO coordinates workflows across systems for training and simulation.
The Groot model learns from human demonstrations to assist in tasks and emulate human movement.
Nvidia's technologies allow for understanding humans from videos, training in simulations, and deploying to physical robots.
Jetson Thor robotics chips provide the intelligence for the future of AI-powered robotics.
The advancement in generative AI technology is leading to the creation of new infrastructure dedicated to specialized tasks.
02:01:31
This new industrial revolution is centered around generative AI with trillion parameters, enabling the development of incredibly valuable software.
The future of computing involves the creation of new types of software that can be easily distributed and accessed in the cloud or on-the-go.
Personal intelligence can be encapsulated in a portable format known as Nims.
Nvidia is focusing on developing proprietary applications and chatbots using AI technology and infrastructure to support the robotic systems of the future.