A paper released today describes in the greatest detail to date the atmospheres on distant planets.
Seeking the origins of what’s in and beyond the Milky Way, researchers surveyed 25 exoplanets, bodies that orbit stars far beyond our solar system. Specifically, they studied hot Jupiters, the largest and thus easiest to detect exoplanets, many sweltering at temperatures over 3,000 degrees Fahrenheit.
Their analysis of these torrid atmospheres used high performance computing with NVIDIA GPUs to advance understanding of all planets, including our own.
Hot Jupiters Shine New Lights
Hot Jupiters “offer an incredible opportunity to study physics in environmental conditions nearly impossible to reproduce on Earth,” said Quentin Changeat, lead author of the paper and a research fellow at University College London (UCL).
By analyzing trends across a large group of exoplanets, they shine new light on big questions.
“This work can help make better models of how the Earth and other planets came to be,” said Ahmed F. Al-Refaie, a co-author of the paper and head of numerical methods at the UCL Centre for Space Exochemistry Data.
Parsing Hubble’s Big Data
They used the most data ever employed in a survey of exoplanets — 1,000 hours of archival observations, mainly from the Hubble Space Telescope.
The hardest and, for Changeat, the most fascinating part of the process was determining what small set of models to run in a consistent way against data from all 25 exoplanets to get the most reliable and revealing results.
“There was an amazing period of exploration — I was finding all kinds of sometimes weird solutions — but it was really fast to get the answers using NVIDIA GPUs,” he said.
Millions of Calculations
Their overall results required heady math. Each of about 20 models had to run 250,000 times for all 25 exoplanets.
“I expected the A100s might be double the performance of V100s and P100s I used previously, but honestly it was like an order of magnitude difference,” said Al-Refaie.
Orders of Magnitude Gains
A single A100 GPU gave a 200x performance boost compared to a CPU.
Packing 32 processes on each GPU, the team got the equivalent of a 6,400x speedup compared to a CPU. Each node on Wilkes3 delivered with its four A100s the equivalent of up to 25,600 CPU cores, he said.
The speedups are high because their application is amazingly parallel. It simulates on GPUs how hundreds of thousands of light wavelengths would travel through an exoplanet’s atmosphere
On A100s, their models complete in minutes work that would require weeks on CPUs.
The GPUs ran the complex physics models so fast that their bottleneck became a CPU-based system handling a much simpler task of determining statistically where to explore next.
“It was a little funny, and somewhat astounding, that simulating the atmosphere was not the hard part — that gave us an ability to really see what was in the data,” he said.
A Wealth of Software
Al-Refaie employed CUDA profilers to optimize jobs, PyCUDA to optimize the team’s code and cuBlas to speed up some math routines.
“With all the NVIDIA software available, there’s a wealth of things you can exploit, so the team is starting to spit out papers quickly now because we have the right tools,” he said.
They will need all the help they can get, as the work is poised to get much more challenging.
Getting a Better Telescope
The James Webb Space Telescope comes online in June. Unlike Hubble and all previous instruments, it’s specifically geared to observe exoplanets.
The team is already developing ways to work at higher resolutions to accommodate the expected data. For example, instead of using one-dimensional models, they will use two- or three-dimensional ones and account for more parameters like changes over time.
“If a planet has a storm, for example, we may not be able to see it with current data, but with the next generation data, we think we will,” said Changeat.
The rising tide of data opens a door to apply deep learning, something the group’s AI experts are exploring.
It’s an exciting time, said Changeat, who’s joining the Space Telescope Science Institute in Baltimore as an ESA fellow to work directly with experts and engineers there.
“It’s really fun working with experts from many fields. We had space observers, data analysts, machine-learning and software experts on this team — that’s what made this paper possible,” Changeat said.
Learn more about the paper here.
Image at top courtesy of ESA/Hubble, N. Bartmann