GTC China (early) Jensen Keynote
Category: China Keynote · Year: 2018 · ▶ Watch
Segments (11)
- 00:00 · Introduction & Vision
- Opening video and Jensen Huang’s introduction to the record-breaking GTC China event.
- 03:50 · Reinventing Computer Graphics
- Introduction of the Turing architecture, RT Cores, Tensor Cores, and real-time ray tracing.
- 17:10 · Deep Learning Super Sampling (DLSS)
- Explanation of how Tensor Cores enable DLSS to improve image quality and performance.
- 27:00 · The Path Forward for Computing
- Discussion on the end of Moore’s Law and the need for accelerated computing across the full stack.
- 44:20 · AI Transforming Computing
- How AI is changing the landscape of High Performance Computing (HPC) and Hyperscale data centers.
- 50:30 · NVIDIA HGX-2
- Introduction of the HGX-2 platform for large-scale HPC and AI training.
- 58:00 · NVIDIA T4 Cloud GPU & TensorRT
- Launch of the T4 GPU and TensorRT Inference Server for hyperscale AI inference.
- 78:10 · RAPIDS: Accelerating Data Science
- Introduction of the RAPIDS open-source libraries for GPU-accelerated machine learning and data analytics.
- 97:40 · AI Automating the World
- Transition to autonomous machines, robotics, and the introduction of the Xavier SoC.
- 100:40 · NVIDIA AGX Platforms
- Overview of domain-specific AI platforms: Clara (Medical), Metropolis (Cities), Isaac (Robotics), and DRIVE (Auto).
- 107:20 · NVIDIA DRIVE & Autonomous Vehicles
- Updates on the DRIVE AV platform and announcements of partnerships with Chinese automakers and mobility services.
Product Announcements (7)
- [03:50] Turing Architecture
- New GPU architecture featuring RT Cores for ray tracing and Tensor Cores for AI.
- specs: 10 Giga Rays, 14 TFLOPS, 114 TFLOPS Tensor Core
- availability: N/A
- [20:50] RTX 2070
- Consumer graphics card based on the Turing architecture.
- specs: Faster than Pascal 1080 Ti
- availability: $499
- [50:30] NVIDIA HGX-2
- AI and Data Analytics platform for HPC.
- specs: 16 V100 GPUs, 2 PFLOPS, 512GB HBM2 memory, 80,000 CUDA cores
- availability: N/A
- [58:00] NVIDIA T4 Cloud GPU
- Universal GPU for hyperscale data centers, optimized for inference.
- specs: 70W, 65 TFLOPS FP16, 130 TOPS INT8, 260 TOPS INT4
- availability: N/A
- [61:30] TensorRT Inference Server
- Open-source containerized inference microservice.
- specs: Supports multiple frameworks, runs on Kubernetes, maximizes GPU utilization
- availability: Available now
- [78:10] RAPIDS
- Open-source suite of data processing and machine learning libraries.
- specs: cuDF (Pandas-like), cuML (Scikit-Learn-like), cuGraph, built on Apache Arrow
- availability: Open source
- [98:10] Xavier
- World’s first AI computing processor designed for autonomous machines.
- specs: 9 Billion transistors, 350mm2, 12nFFN, 30 TOPS, 30W
- availability: In full production
Specific Numbers (6)
| Timestamp | Metric | Value | Context |
|---|---|---|---|
| 08:30 | TFLOPS | 114 | Turing architecture Tensor Core performance. |
| 20:50 | $ | 499 | Price of the RTX 2070 GPU. |
| 35:10 | Supercomputers | 127 | Number of NVIDIA GPU systems on the TOP500 list. |
| 50:30 | PFLOPS | 2 | Performance of the HGX-2 platform. |
| 59:30 | TOPS | 260 | INT4 performance of the T4 Cloud GPU. |
| 98:10 | Transistors | 9 Billion | Number of transistors in the Xavier SoC. |
Benchmark Claims (3)
- [18:50] Ray Tracing + DLSS Performance: RTX 2070
- vs: GTX 1080 Ti
- gain: 3.5x performance improvement.
- [58:10] Inference Performance: T4 GPU
- vs: CPU
- gain: Up to 40x faster inference performance.
- [87:40] Data Science Workflow (ETL + ML): 1 DGX-2
- vs: 20 CPU Nodes
- gain: Reduced processing time from hours to minutes (50x speedup).
Customer Stories (5)
- [09:00] NetEase
- Integrated RTX ray tracing into their game ‘Justice’.
- outcome: Created the first China RTX game with realistic reflections and lighting.
- [19:20] Kingsoft
- Integrated DLSS into their game ‘JX3’.
- outcome: Created the first China DLSS-accelerated game, improving performance by 1.8x.
- [91:00] BGI
- Used RAPIDS XGBoost for cancer immunotherapy and normal cell classification.
- outcome: Achieved 10x faster processing compared to CPU.
- [105:00] JD.com, Meituan, Cainiao
- Selected Jetson AGX Xavier for their autonomous delivery robots.
- outcome: Enabled complex AI processing for last-mile delivery logistics.
- [118:20] FAW, Full Truck Alliance, Plus.ai
- Partnered to develop an autonomous truck hailing service using NVIDIA DRIVE.
- outcome: Targeting production in 2021 to solve driver shortages and reduce costs.
Key Technologies (4)
- RT Core: Dedicated hardware on Turing GPUs that accelerates real-time ray tracing calculations.
- DLSS (Deep Learning Super Sampling): Uses Tensor Cores and deep learning to generate high-resolution images from lower-resolution inputs, improving frame rates.
- TensorRT Inference Server: A containerized microservice that maximizes GPU utilization by running multiple models concurrently.
- RAPIDS: Open-source libraries that allow data scientists to execute end-to-end data science and analytics pipelines entirely on GPUs.
Demos Shown (6)
- [09:40] Real-time ray tracing demo in the game ‘Justice’, showcasing reflections and lighting.
- True
- [21:50] Porsche 911 Speedster Concept rendered in real-time using Turing.
- True
- [65:00] T4 Inference Server classifying thousands of flower images per second in real-time.
- True
- [86:20] RAPIDS accelerating a mortgage risk analysis workflow from hours on CPUs to minutes on GPUs.
- True
- [111:30] NVIDIA DRIVE AV software stack navigating a car autonomously on a highway.
- True
- [123:00] Project Sol cinematic demo rendered in real-time, showcasing advanced robotics and graphics.
- True
Predictions / Commitments (3)
- [27:00, Current and ongoing] Moore’s Law has come to an end, requiring a full-stack optimization approach to continue advancing computing performance.
- [38:10, Current and ongoing] AI is automating the world, impacting industries from healthcare to transportation and manufacturing.
- [97:40, Future] In the future, AI will empower all moving machines, leading to billions of autonomous robots and vehicles.
Companies Mentioned (3)
Baidu, Tencent, Alibaba, Huawei, Inspur, Lenovo, Sugon, QCT, Supermicro · Baidu Cloud, Tencent, JD Cloud, iFLYTEK · Volvo Cars
Notable Quotes (2)
Moore’s Law has come to an end. — Jensen Huang @ 27:00
The more you buy, the more you save. — Jensen Huang @ 90:40
Key Topics
Turing Architecture · Ray Tracing · DLSS · High Performance Computing (HPC) · Hyperscale Data Centers · T4 GPU · TensorRT · RAPIDS · Data Science · Xavier SoC · Autonomous Machines · NVIDIA DRIVE · Autonomous Vehicles
Takeaways
- NVIDIA is reinventing computer graphics with the Turing architecture, combining ray tracing and AI (DLSS) for photorealistic real-time rendering.
- With the end of Moore’s Law, NVIDIA advocates for a full-stack accelerated computing approach to continue driving performance gains.
- The HGX-2 platform unifies HPC and AI workloads, providing massive computational power for complex simulations and deep learning.
- The T4 Cloud GPU and TensorRT Inference Server are designed to efficiently scale AI inference in hyperscale data centers.
- RAPIDS brings GPU acceleration to traditional data science and machine learning workflows, drastically reducing processing times.
- The Xavier SoC and AGX platforms are enabling a new era of autonomous machines across robotics, healthcare, smart cities, and transportation.
- NVIDIA DRIVE is seeing broad adoption among automakers, EV startups, and mobility services for developing autonomous driving solutions.