Perlmutter, one of the fastest supercomputers, went online yesterday at Scientific Computing Center for National Energy Research In California. The supercomputer was developed specifically for artificial intelligence.
According to Nvidia, the Perlmutter system is the fastest in the world when it comes to handling 16- and 32-bit mixed-stage math used in AI applications. “The supercomputer will deal with some of the most difficult scientific challenges in astrophysics and climate science, such as creating a three-dimensional map of the universe and exploring subatomic interactions of green energy sources.” The supercomputer is powered by no less than 6159 Nvidia A100 Tensor Core GPUs. Tensor Core is the most advanced GPU developed by Nvidia to date. This computing power makes Perlmutter the largest A100 graphics processing system in the world, according to Nvidia, which can process nearly a trillion floating-point operations per second.
The largest 3D map in the universe
“We are in the exascale era of AI,” said Dion Harris, Nvidia’s chief product marketing manager. In a blog post, he explains that researchers will use Perlmutter to create the largest 3D map of the universe ever. For this, the system will use data from the Dark Energy Spectroscopy Tool (DESI). The search tool can capture images of up to 5000 universe with a single exposure. Harris explains that Perlmutter will analyze data from dozens of DESI exposures every night. This is to help astronomers determine where to point the device the next day. Using a 3D map of the universe, researchers can learn more about dark matter, which is said to be responsible for the accelerating expansion of the universe.
Perlmutter Atomic Simulation
NERCS plans to make the Perlmutter system available to more than 7,000 researchers worldwide. The supercomputer will also be used to investigate reactions at the atomic level, for example to develop better batteries and new biofuels. Harris points out that “traditional supercomputers can barely handle the mathematics required to generate an atomic simulation of a few nanoseconds.” “But by combining high-fidelity simulations with machine learning, scientists can study atoms over a longer period of time.” A100 Tensor Cores would be uniquely suited for this, as it can accelerate both dual-precision floating point calculations for simulations and mixed-precision calculations.