GTC DC 2019 (Ian Buck Keynote)

Category: DC Keynote · Year: 2019 · ▶ Watch

Switch language → 中文

Segments (10)

  • 00:00 · Introduction and Event Overview
    • Ian Buck welcomes attendees to GTC DC, highlighting record attendance and the growth of the East Coast AI hub.
  • 03:45 · The Big Bang of Modern AI
    • A review of AI’s evolution from basic image classification to advanced capabilities like segmentation, language understanding, and generative models.
  • 11:45 · Data Processing and RAPIDS
    • Discussion on the challenges of data preparation and how the RAPIDS software stack accelerates ETL workflows using GPUs.
  • 17:15 · AI Training and Supercomputing
    • Addressing the exponential growth in neural network complexity and the introduction of the DGX SuperPOD to handle massive training workloads.
  • 23:00 · Inference and Conversational AI
    • Exploring the shift towards complex, real-time inference pipelines required for conversational AI and smart devices.
  • 29:25 · Applying AI to Healthcare
    • How the Clara SDK and Transfer Learning Toolkit are helping the medical industry overcome data and talent shortages to deploy AI.
  • 38:05 · 5G, Smart Cities, and Edge Computing
    • The intersection of 5G networks and AI, featuring the announcement of the EGX Edge Supercomputing Platform and real-world deployments.
  • 50:30 · Autonomous Vehicles
    • An overview of the end-to-end workflow and in-car perception stack required to develop and deploy self-driving cars.
  • 56:25 · Robotics and the Isaac SDK
    • How AI is moving robots from repetitive tasks in cages to collaborative, intelligent machines operating in dynamic environments.
  • 64:20 · Global AI Strategies and Workforce Training
    • Concluding remarks on the economic impact of AI, national strategies, and NVIDIA’s efforts to train the next generation of AI developers.

Product Announcements (3)

  • [43:10] NVIDIA EGX Edge Supercomputing Platform
    • A scalable server platform designed to process AI workloads at the edge.
    • specs: Supports 1 to 4 GPUs, cloud-native architecture, supported by major OEMs like Dell, HP, and Lenovo.
    • availability: Available through OEM partners.
  • [41:15] NVIDIA Aerial
    • A software developer kit for building 5G virtual radio access networks (vRAN).
    • specs: GPU-accelerated 5G stack to enable edge computing on telecommunications infrastructure.
    • availability: Not explicitly stated.
  • [19:05] Project Megatron
    • A massive natural language processing model developed by NVIDIA Research.
    • specs: 8.3 billion parameters, 24x larger than Google’s BERT model.
    • availability: Open-source research project.

Specific Numbers (10)

Timestamp Metric Value Context
02:50 Event Registrants 3,500 Number of attendees at GTC DC, representing a 30% increase from the previous year.
14:10 Data Volume 1 Terabyte per day Amount of telemetry data processed daily by Charter Spectrum.
19:10 Model Size 8.3 Billion parameters The size of NVIDIA’s Project Megatron NLP model.
20:40 Compute Power 9.4 Petaflops The performance capability of the DGX SuperPOD architecture.
31:45 Workload 8,000 images per day The average number of images a radiology department must process daily.
38:50 Network Density 1 Million devices per square kilometer The device density capability promised by 5G networks.
48:55 Processing Volume 485 Million mail pieces per day The daily sorting volume handled by the United States Postal Service.
60:15 Hardware Cost $99 The price of the Jetson Nano developer kit for edge AI and robotics.
65:30 Economic Impact $13 Trillion McKinsey’s projection for the increase in global economic activity driven by AI by 2030.
68:55 People Trained 180,000 The number of individuals trained by the NVIDIA Deep Learning Institute.

Benchmark Claims (4)

  • [14:55] Data ETL Processing: 4 minutes
    • vs: 8 days on CPU
    • gain: Massive reduction in processing time using RAPIDS on GPUs.
  • [20:55] ResNet-50 Training: 80 seconds
    • vs: Previous generation hardware
    • gain: Demonstrates the extreme scaling capability of the DGX SuperPOD.
  • [35:50] Medical Model Training: Under 30 minutes with 7,000 images
    • vs: 19 hours with 250,000 images (training from scratch)
    • gain: Drastic reduction in both data requirements and compute time using the Transfer Learning Toolkit.
  • [48:55] Package Sorting Speed: 10x Faster
    • vs: Previous non-AI sorting methods
    • gain: Significant throughput increase for USPS operations.

Customer Stories (4)

  • [13:55] Charter Spectrum
    • Implemented NVIDIA RAPIDS to process network telemetry data from 500,000 access points.
    • outcome: Reduced data processing time from 8 days on CPUs to just 4 minutes on GPUs.
  • [45:50] City of Dubuque, Iowa
    • Deployed DeepStream and EGX servers to analyze feeds from 200 traffic cameras.
    • outcome: Enabled real-time anomaly detection, such as identifying vehicles driving the wrong way on highways.
  • [48:45] United States Postal Service (USPS)
    • Adopted NVIDIA AI, DeepStream, and EGX servers for package sorting across 200 facilities.
    • outcome: Achieved 10x faster sorting speeds and higher accuracy compared to previous methods.
  • [58:35] John Deere
    • Developed an AI-powered agricultural robot that uses computer vision to distinguish between crops and weeds.
    • outcome: Reduced pesticide usage by up to 90% by only spraying identified weeds.

Key Technologies (6)

  • RAPIDS: An open-source suite of data processing and machine learning libraries that execute entirely on GPUs.
  • DGX SuperPOD: A scalable, high-performance AI supercomputing infrastructure designed for massive model training.
  • Clara SDK: An application framework tailored for healthcare, enabling AI-assisted annotation and medical image processing.
  • Transfer Learning Toolkit: A tool that allows developers to adapt pre-trained neural networks to new, specific datasets with minimal data and time.
  • DeepStream: A streaming analytics toolkit for AI-based multi-sensor processing, video, and image understanding.
  • Isaac SDK: A software framework and simulation environment for developing, training, and deploying AI in robotics.

Demos Shown (6)

  • [07:10] GauGAN demo where simple painted shapes are instantly converted into photorealistic landscapes.
    • True
  • [08:25] 3D Pose Estimation tracking a person’s movements to animate a virtual astronaut on the moon.
    • True
  • [09:05] Reinforcement learning simulation showing a virtual character learning how to perform a backflip.
    • True
  • [27:20] Audio playback comparing a human voice to an AI-generated voice using WaveNet.
    • True
  • [52:55] In-car video demonstrating the DRIVE perception stack identifying lanes, vehicles, pedestrians, and traffic lights in real-time.
    • True
  • [61:00] Isaac simulation showing virtual robot arms learning to pick up and stack blocks.
    • True

Predictions / Commitments (5)

  • [23:45, By next year (2020)] Half of the world’s internet searches are going to be done by voice.
  • [24:08, By 2022] Over 50% of households in the US will own a smart speaker.
  • [34:40, Next year] Half of the world’s internet searches will be done by voice.
  • [59:40, Not explicitly dated, but framed as a near-future trend.] The home food delivery market (ghost restaurants) is projected to reach a hundred billion dollars.
  • [65:30, By 2030] AI is going to deliver or increase global economic activity by 13 trillion dollars.

Companies Mentioned (5)

OpenAI · Google · AWS, Google Cloud, Microsoft Azure · Microsoft · Dell, HP, Lenovo, Supermicro

Notable Quotes (4)

It is great to see DC becoming basically an East Coast hub for AI innovation. — Ian Buck @ 03:18

Basically data scientists are going to build the fastest, biggest, baddest networks they can, as long as they can train in about a week. — Ian Buck @ 17:23

If I can’t run all those neural networks and do all that search and get the audio back in a third of a second, the user experience just falls off a cliff. — Ian Buck @ 28:44

It’s so exciting that people don’t know why it’s exciting. It actually is exciting. — Ian Buck @ 38:18

Key Topics

Artificial Intelligence · GPU Acceleration · Data Science · Natural Language Processing · Conversational AI · Healthcare AI · Edge Computing · 5G Networks · Autonomous Vehicles · Robotics · Smart Cities · Transfer Learning

Takeaways

  • AI capabilities have expanded rapidly from simple image classification to complex, multi-modal tasks like conversational AI and autonomous navigation.
  • Data preparation is a significant bottleneck in AI workflows; GPU-accelerated tools like RAPIDS can reduce ETL times from days to minutes.
  • The exponential growth in the size of AI models, particularly in NLP, necessitates massive, scalable supercomputing infrastructure like the DGX SuperPOD.
  • To meet the low-latency demands of modern applications (e.g., smart cities, 5G, robotics), AI inference must move from centralized clouds to edge computing platforms like EGX.
  • NVIDIA is facilitating industry-specific AI adoption by providing tailored software frameworks (Clara, Metropolis, DRIVE, Isaac) that lower the barrier to entry.
  • Transfer learning is a critical technique for applying AI in specialized domains, allowing organizations to achieve high accuracy with significantly less data and training time.