Recently, astronomers and astrophysicists began to turn to the computing performance of GPU and AI, trying to get more information from the images taken by telescopes.
Research teams from the University of California, Santa Cruz and Princeton University have been challenging the limits of this field. Led by Brant Robertson of the University of California, Santa Cruz and Evan Schneider, winner of NASA Hubble Scholarship, the team focused on optimizing the use of NVIDIA GPU and deep learning tools to meet the requirements of large-scale computing.
Their goal is to expand its computational performance and conduct more accurate hydrodynamic simulation, so as to further understand the formation of galaxies.
The team shifted the workload from CPU to GPU from the beginning. In this way, they can measure the substances inside and outside the surface of three-dimensional grid cells, just as they can solve multiple Rubik's cubes at the same time.
With the help of CUDA platform, the research team can transfer a series of grids to GPU for necessary calculations, so as to obtain more detailed simulation results.
After releasing most of the performance of the system, they turned their attention to the more powerful NVIDIA GPU cluster device, namely the Titan supercomputer located in Oak Ridge National Laboratory of the US Department of Energy. However, in order to carry out higher resolution simulation, they need more powerful code to control the powerful performance of Titan's more than 65,438+06,000 Tesla GPUs.
Schneider is the best person for this task. She was a student of Robertson when she was a graduate student, and now she is a postdoctoral researcher at Princeton University. Schneider wrote a code of fluid dynamics accelerated by GPU, named CHOLLA, namely "Computational Fluid Dynamics on Parallel Architecture".
"How did the Milky Way wind get there? What is the cause of galaxy wind? How does the galactic wind control the quality of galaxies? These are all questions we want to study, but they are very difficult to calculate, "Robertson said. "Evan is the first person who can solve this problem accurately."
Using the CHOLLA code written several years ago, Schneider and Robertson can execute 1 100 million core hours on Titan. The uniqueness of this code is that it can perform all operations on the GPU, which enables the research team to conduct complex simulations of NVIDIA DGX and DGX- 1 deep learning systems in their laboratories, and then transmit the simulation results to Titan for expansion.
"You want to take advantage of the floating-point computing performance of GPUs, and you don't want to spend time waiting for information to pass back and forth between GPUs," Robertson said. "Spend as much time as possible on the GPU, which is what you want to achieve."
CHOLLA can expand a large number of GPUs, enabling research teams to test and calculate 550 billion cells, which Robertson called "one of the largest hydrodynamic simulations in astrophysics".
Another student, Ryan Hausen, developed a deep learning framework called Morpheus, which uses original telescope data to classify galaxies, paving the way for more ambitious research projects. With this framework, they are expected to process the large-scale measurement results of billions of galaxies on DGX system.
In addition, another plan is also planned-Robertson hopes to compute on Summit, the world's most powerful supercomputer powered by Volta GPU in NVIDIA. He believes that with the expansion of GPU memory by Summit, CHOLLA can make them achieve greater research results than using Titan.
"The computing performance of NVIDIA's GPU enables us to perform numerical simulations that were impossible in the past," Robertson said. "Next, we intend to use NVIDIA GPU to challenge more possibilities."