RTX: BRINGING AI & ADVANCED GRAPHICS TO VISUAL COMPUTING RAJ MIRPURI – VP PROFESSIONAL VISUALIZATION NVIDIA ACCELERATES ALL WORKLOADS
SCIENTIFIC DEEP RENDERING MACHINE RESEARCH LEARNING & VIZ LEARNING
CUDA
GPU
2 VISUAL COMPUTING CHALLENGES TODAY
Increased Complexity Increased Demand for Massive Data, & Quality Data Analysis Real-time Insight
Need for Mixed Visualization & Compute
Rendering
3 NVIDIA POWERS THE NEXT ERA OF COMPUTING
Turing RTX Architecture RTX Platform
Turing SM Display
RT Cores NVLINK
Tensor Cores Video Encode/Decode
Massive Memory
Accelerated offline rendering Data Science platform AI at the Edge Powerful virtual workstations
4 RTX SERVER ENABLES AI FOR VISUAL COMPUTING
Rendering Data Science AI at the Edge
Accelerate offline batch Enable multiple GPU configurations to Unlock benefits as bandwidth, rendering with the power of develop on larger data sizes to find latency and availability to resources multi-GPU acceleration business impacting and changing results are not ubiquitous and plentiful
Virtualization and Mixed Workloads
Empower designers and artists anywhere in the world with the high end RTX GPUs
5 RTX AI FOR RENDERING AN EXPLOSION OF RICH CONTENT
Annual Spend on Original Content $15B Amazon Netflix $10B
$5B
$0B 2017 2018 2019 2020 2021 2022
1993: Jurassic Park had a total of 63 VFX shots taking The shot had over 8 million instanced lights, which a year to create initially took 1,000 hours to render each frame
2018: Marvel’s Avengers: End Game had more than By the end of production, it still took 50 hours to 3,000 shots render each fully optimized frame
7 CREATE MORE, WAIT LESS WITH QUADRO
Note: CPU = i9-7900X; GPU = NVIDIA RTX GPU; video playback at 2x speed 8 EXPONENTIAL POWER AT 1/4 THE COST
240 Dual 12-core Skylake CPU Servers 4 RTX 8-GPU Servers 144 kW 13 kW $2M Render Farm $500,000 1/4 the Cost 1/10 the Space 1/11 the Power
Interactive Content Creation Batch Rendering Interactive RTX-Accelerated RenderingBatch
9 25X ACCELERATED RENDERING FOR NETFLIX
CPU Node RTX Server Improvement (Dual Skylake) (4 x RTX 8000) Render time 38 min 6 min 6x (1 frame) Total render time 76 hours 3 hours 25x (120 frames)
# of nodes 25 1 25x
Power (kW) 13.2 1.9 7x
NETFLIX Lost In Space scene: renders in a fraction Acquisition cost $188k $28k 7x of the time using RTX Server Cost of power (5 • 6x faster for a single frame $68k $10k 7x • 25x faster for the entire shot yrs) Total cost $256k $38k 7x
10 RTX FOR DESIGN AND RENDERING Accelerated Workflows with RTX
DESIGN RENDERING
OpenGL DirectX Vulkan OptiX IndeX VXGI VRWorks PhysX MDL vMaterials GVDB DESIGNWORKS NvPro Video Codec SDK
RTX SERVER
DESKTOP | MOBILE WORKSTATION DATACENTER CLOUD RTX AI FOR DATA SCIENCE 3M DATA SCIENTISTS AT WORK TODAY
CONSUMER INTERNET OIL & GAS Ad Personalization Sensor Data Tag Mapping Click Through Rate Optimization Anomaly Detection Churn Reduction Robust Fault Prediction
FINANCIAL SERVICES MANUFACTURING Claim Fraud Remaining Useful Life Estimation Customer Service Chatbots/Routing Failure Prediction Risk Evaluation Demand Forecasting
HEALTHCARE TELECOM Improve Clinical Care Detect Network/Security Anomalies Drive Operational Efficiency Forecasting Network Performance Speed Up Drug Discovery Network Resource Optimization (SON)
RETAIL AUTOMOTIVE Supply Chain & Inventory Mgmt Intelligent Customer Interactions Price Mgmt / Markdown Optimization Connected Vehicle Predictive Maintenance Promotion Prioritization And Ad Targeting Forecasting, Demand, & Capacity Planning
13 RTX STREAMLINES DATA SCIENCE WORKFLOWS
ANOTHER…
@*#! Forgot to Add Same Number of Iterations Train Model GET A COFFEE a Feature in Much Less Time Validate Restart Data Prep Start Data Prep Test Model Start Workflow Workflow 12 GET A COFFEE 12 Experiment with GET A COFFEE Optimizations and Repeat Switch to Decaf Configure Data Prep Workflow CPU- GPU- 9 POWERED 3 9 POWERED 3 WORKFLOW WORKFLOW
Find Unexpected Null Values Stored as String… Dataset 6 Dataset 6 Downloads Restart Data Prep Downloads Overnight Workflow Again Overnight
Stay Late Go Home on Time
Dataset Collection Analysis Data Prep Train Inference
14 10X FASTER PERFORMANCE WITH RTX
Seconds (lower is better)
~ 30X ~ 8X ~ 10X Data Prep Faster Training w/XGBoost Faster End-to-End Faster than than than 300 200 600 CPU CPU CPU 180 250 500 160
140 200 400 120
150 100 300
80 100 200 60
40 50 100 20
0 0 0 CPU 1x RTX8000 2x RTX8000 CPU 1x RTX8000 2x RTX8000 CPU 1x RTX8000 2x RTX8000
15 End-to-end time = Data Prep + Conversion + Training + Validation Dataset: Mortgage Data 2015-2016 CPU: dual Gold [email protected] 3.7GHz Turbo (Skylake), ETL with Dask + Pandas 20X SPEEDUP FOR ARUP ANALYTICS
Opportunity: 200 Data Scientist
Use Case: Distributed infrastructure assets, precipitation risk assessment
Data Preparation: 50,000 CSV files to extract data and then aggregate in multiple formats
Model Training: Open source python for analytics, pandas, GeoPandas, Shapely, NumPy
“The NVIDIA-powered Data Science Workstation[’s]… combination of well-designed software and highly performant hardware provides a 20x and higher speed-ups in our analytics work and our team found its ease of use liberating.”
Steve Walker, Associate Director, Arup, Advanced Digital Engineering
16 RTX FOR DATA SCIENCE & AI From Data Science to NVIDIA Accelerated Data Science with CUDA-X AI
17 RTX AI AT THE EDGE SMART AND SAFE CITIES NEED AI
Image Classification 1,000M 97% 800M Human
600M Deep Learning 400M
Accuracy 74% 200M Hand-coded CV
0M 2016 2020 2010 2011 2012 2013 2014 2015 2016
1 billion installed security Challenging real world conditions AI achieves super human results cameras WW (2020) Traditional video analytics not AI driven intelligent video analytics 30 billion frames per day trustworthy
19 RTX ENABLES AI AT THE EDGE
EXTRACTING NEW VALUE SAFETY COMPLIANCE RETAIL STORE VIDEO SECURITY & PUBLIC SAFETY
20 21 REAL-TIME INSIGHTS TO FIND RESOLUTION
Finding lost people Capturing a criminal Adding 100 virtual Improving traffic Reducing peak traffic in 30 minutes before he strikes again security guards safety congestion by 15%
22 RTX AI FOR MIXED WORKLOADS WORKFLOWS DEMAND MIXED WORKLOADS
VISUALIZATION COMPUTE
Windows 10 2D EDA 3D Apps Deep Learning HPC
24 MIXED WORKLOADS ARE EVERYWHERE Professional workflows are getting more demanding
Manufacturing Architecture Medical Imaging Energy
Media & Entertainment Automotive Finance Defense COMPUTER AIDED ENGINEERING WORKFLOW Bottlenecks Impact Design Iterations and Time to Market
CAD
Pre-Processing (CAE)
HPC Solver (CAE)
Post-Processing (CAE) Addressing the Bottlenecks of FEA Simulation Source: Tech-Clarity 2017
26 SIMULATION WITHIN DESIGN POWERED BY RTX
27 QUADRO VIRTUAL WORKSTATION STREAMS FROM RTX SERVER To Any Device, Anywhere
NVIDIA Quadro Virtual Data Center Workstation Software • Now supports RTX 6000 and RTX 8000 • Brings world’s most powerful virtual workstation to RTX Server platform • Flexibly provision virtual workstations or a combination of virtual workstations and render nodes from a single RTX Server • Extends power of RTX platform to designers, on any device, anywhere
28 QUADRO RTX VIRTUAL WORKSTATIONS
Type of User Light Users Medium Users Heavy Users
Recommended NVIDIA T4 or P6 NVIDIA T4 or P6 NVIDIA Quadro RTX 8000, Solution Quadro Virtual Data Quadro vDWS RTX 6000, P40 or V100 Center Workstation with Quadro vDWS
GPU Memory 16 GB 16 GB 48 GB/32 GB/24 GB
Equivalent Multiple Quadro P1000 Up to Quadro P4000 Up to Quadro RTX 8000 Performance
Replaces K2, M60, P4, M6 K2, M60, P4, M6 N/A
NEW TO VDI
29 RTX POWERS AI FOR VISUAL COMPUTING
Accelerate offline batch rendering with the power of multi-GPU acceleration
Enable multiple GPU configurations to develop on larger data sizes to find business impacting and changing results
Unlock benefits as bandwidth, latency and availability to resources are not ubiquitous and plentiful
Empower designers and artists anywhere in the world with the high end RTX GPUs
30