7. Next Level of AI GPGPU
in Space Applications
Aitech’s S-A1760 Venus™: most
powerful and smallest space AI GPGPU in
small form factor (SFF). Suitable for the next
gen of short duration spaceflight, NEO and
LEO.
7
8. Accelerated prediction and real-time detection
SATELLITE
IMAGERY
WEATHER
TERRAIN
Infrequent updates
Minutes/Hours
Hours/Days
INCOMING
SENSOR
DATA
WILDFIRE
PREDICTION
MODEL
OMNIVERSE
REGIONAL DIGITAL
TWIN
LIST OF COARSE
PROBABLE FIRE
LOCATIONS
OBSERVATIONS
EVENT
PREDICTION
THIRD PARTY
PLANNING APPS/OV
CLIENTS
Fire front
simulations
Detected, refined fire
locations and attributes
“State model updates”
DATA CENTER
EDGE
• Coarse
resolution
• Broad area
coverage
• Regular updates
• Acceleration
opportunity
• High resolution collection – ‘soda straw’
• Potentially comms-challenged
Either published out or
offline loaded prior to
mission
Physics-informed AI
Model?
*currently, we
are working
here
Accelerated fired
detection and
orthorectification
Edge processing as fast as camera
runs
• Precision mapping onto terrain
• Critical fire location / attributes
9. 9
“NVIDIA is proud to have been one of ‘the
Mavericks’ at the inception of the Frontier
Development Laboratory. Our headquarters are
named Endeavor and Voyager – mighty ships for
our journey to the stars. We are delighted to
support FDL as we push the frontiers for the
human race.”
JENSEN HUANG, Founder, President & CEO,
NVIDIA Corporation
13. 13
Neural Radiance Fields (NeRFs) introduced in 2020
NeRFs are an astonishing new method for capturing reality
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. 2020.
14. 14
Text Input for Latent Space Embedding
Instant-NGP NeRFs + Hashgrid Differentiable Rendering
Pure Noise
Diffusion Process
Final Image
Diffusion Models
22. Physics Simulation in Omniverse
Multiple Ways to Ingest and Simulate Physics in Omniverse
NVIDIA PhysX
Rigid & Soft Body Dynamics, Destruction, Fluid & Fire
Import Physics Instance
Offline from External 3rd Party Application
NVIDIA Modulus
Accelerated with Physics-ML Platform
23. DTE USE CASES
Already working with NOAA on a sea-surface system – land, atmosphere, space weather & cryo (ice) coming soon
EXTREME WEATHER
Predict weather. Detect storms
ENERGY
Downscale. Predict energy. Optimize.
WILDFIRE
Satellite detection, fire prediction,
smoke forecasting
OCEAN AND ICE
Predict ice-melt, sea-level, ecosystem
changes
24.
25. Escalating adventures in using machine learning
To parameterize convection in climate models, and new perspectives from autoregressive weather forecasting in industry
Mike Pritchard, Associate Professor at UC Irvine & Director of Climate Simulation Research at NVIDIA.
26. The low cloud problem
Projections for 2100 wildly disagree
How this system responds to climate change is a multi-trillion dollar question
https://fyfluiddynamics.com/2019/06/urban-centers-during-hurricanes/
27. • Global Storm Resolving
Models can now
capture deep clouds.
• The largest scales of
vertical mixing.
• Grid requirements:
dx ~ 1,000m, dt ~ 20s
Simulating these systems globally requires unfathomable compute
At least 10,000x more than the world’s most computationally demanding global storm-resolving models
• Shallow clouds are way
harder to simulate
• Fine-scale eddies
• Grid requirements: dx ~
100-m, dt ~ 1s
• Tens of thousands of
times more compute-
intensive
Deep Cumulonimbus
Resolvable with today’s global storm resolving tech
Marine Stratocumulus
Computationally out of reach
28. Lack of sub-km physics in climate models should keep us up at night
Boundary layer turbulence
Where humans live
Low cloud formation
Aerosol-cloud interaction
Planetary mixing & climate dynamics
Between boundary & free troposphere
The computational problem matters
29. High resolution climate prediction is a computational challenge
100km
10 km
1 km
100 m
1980 1990 2000 2010 2020 2030 2040 2050 2060
25 km at 10 min
1 km at 20 s (30,000x COMPUTE)
100m at 0.1 s (500 MILLION X COMPUTE)
STORM-RESOLVING
LOW
CLOUD-RESOLVING
Horizontal
grid
spacing
Figure adapted from: Schneider, T., Teixeira, J., Bretherton, C. et al. “Climate goals and computing the future of clouds”. Nature Climate Change 7, 3–5 (2017)
30. • MMF: Multiscale Modeling Framework
• Thousands of “micro-models” of fast, fine-scale physics.
• Embedded in a host global climate model.
• Computationally approachable & scalable.
• Explicit eddies emerging from appropriate equations.
• Also known as "Cloud SuperParameterization"
• Convenient scale exchange arteries for machine learning.
In academia I try to cope using a multi-scale simulation algorithm: “MMF”
The Multi-Scale Modeling Framework approach to global climate modeling
Cloud superparameterization
A multi-scale modeling framework
31. Methods to regionalize unusually high-resolution MMF have emerged
Unexpectedly minimal grid transition artifacts. With load-balancing, yields cost savings of ~ 3x
Peng, Pritchard et al., JAMES, 2022.
32. Worth finding out to leapfrog Moore’s Law and attain appropriate resolution.
100km
10 km
1 km
100 m
1980 1990 2000 2010 2020 2030 2040 2050 2060
25 km at 10 min
1 km at 20 s (30,000x COMPUTE)
100m at 0.1 s (500 MILLION X COMPUTE)
STORM-RESOLVING
LOW
CLOUD-RESOLVING
Horizontal
grid
spacing
Figure adapted from: Schneider, T., Teixeira, J., Bretherton, C. et al. “Climate goals and computing the future of clouds”. Nature Climate Change 7, 3–5 (2017)
Can we use ML to sneak sub-km
physics into climate models decades
ahead of computational schedule?
35. Deuben & Bauer (2018), 6° , 60x30, 1.8K pixels, MLP
WeatherBench, Rasp et al. (2020). 5.625°, 64x32, 2K pixels, CNN
Weyn et al. (2019), 2.5° N.H only, 72x36, 2.6k pixels, ConvLSTM
DLWP, Weyn et al. (2020). 2°, 16K pixels, Deep CNN on Cubesphere/(2021) ResNet
FourCastNet, Pathak et al. (2022), 0.25°, ~1,000,000 Pixels, ViT+AFNO
GNN, Keisler et al. (2022), 1°, 64,000 Pixels, Graph Neural Networks
FourCastNet: NV’s DDWP, first to be trained at ambitious 0.25-deg global resolution
36.
37. Fascinating rate of increase of FourCastNet’s weather prediction skill.
Individual researchers yielding impressive skill gains from ablations independent of large architecture searches
Skill gap reduced by more than half
w.r.t IFS gold standard
Skill gap reduced by more than half
w.r.t IFS gold standard
Acronym Alert:
ACC: Anomaly Correlation Coefficient (metric of weather skill)
IFS: The Integrated Forecast System, a gold standard weather model
FCN: FourCastNet, our digital twin of weather.
Ablations that seem to matter:
• State vector size – adding upper atmosphere helped.
• Only to a limit; adding other things has not immediately helped.
• Model size: # of transformer blocks & hidden dimension.
• Caveat: trades off with inference speed, training cost & memory use.
• Open question: Most efficient route to improving DDWP accuracy.
38. -- Bjorn Stevens, GTC 2022
Tethering via FCN could solve storage, latency crisis for high-res climate prediction
AI nimbly generates details between "checkpoints" saved only infrequently from physics-based climate simulations
40. 40
Lockheed Martin and NVIDIA Selected by NOAA for Earth Observation
Building a Climate Research Data Pipeline with NVIDIA Omniverse
Anomaly
Detection
AI
Algorithms
Data
Fusion
Sensor
Data
Ground
Data
Satellite
Data
BUILD
PHYSICAL WORLD
DIGITAL TWIN
DESIGN OPERATE
SIMULATE
41. 41
EXPLORE OMNIVERSE ENTERPRISE
SEE YOU IN OMNIVERSE
GET ACCESS TO A FREE TRIAL DEVELOP ON OMNIVERSE
DOCUMENTATION
docs.omniverse.nvidia.com
FORUMS
omniverse.nvidia.com/forums
TUTORIALS AND WEBINARS
omniverse.nvidia.com/tutorials
DISCORD
discord.gg/nvidiaomniverse
42. Getting Started With NVIDIA AI
NVIDIA AI Enterprise Trial Programs
Evaluation Software
Requirements: NVIDIA-Certified
System
Free evaluation licenses for on
premises POC
90 days to test and experience
NVIDIA LaunchPad
AI development and deployment trial
program
Deep dive, hands-on labs for AI
practitioners and IT staff
Requires ~8 hrs / Access for 2 wks
43. The Conference for the
Era of AI and the Metaverse
March 20-23, 2023
GTC 2023 Keynote