Portrait of the world’s fastest AI supercomputer: Using 6,159 NVIDIA GPUs

Announced recently at the US National Energy Research Science Computing Center (NERSC), Perlmutter is a supercomputer capable of delivering nearly 4 exaflops of performance. AI for more than 7,000 researchers.

This performance makes Perlmutter the fastest AI system on the planet, using 16- and 32-bit mixed-precision math. And that performance doesn’t even include phase two, due later this year, on the system located at Lawrence Berkeley National Laboratory.

Made up of 6,159 NVIDIA A100 Tensor Core GPUs, it’s the world’s largest A100-powered support system. And now more than 20 applications are ready to become one of the first applications to run on this system. Perlmutter aims to advance the science of astrophysics, climate science and more.

Even in one project, this supercomputer will help assemble the largest 3D map of the visible universe to date. It will process data from the Dark Energy Spectroscopic Instrument (DESI), a type of space camera that can capture up to 5,000 galaxies in a single exposure.

The researchers needed the speed of Perlmutter’s GPU to record dozens of exposures one night, to know where DESI would be the next night. If using previous systems, data preparation will take a whole year or as fast as months, weeks. And Perlmutter will complete the task in just a few days.

“I’m really pleased with the 20x speedup we’ve been getting on the GPU,” said Rollin Thomas, a data architect at NERSC who is helping researchers prepare to use Perlmutter. , in his preparatory work”.

DESI’s map aims to shed light on dark energy, the mysterious physical concept behind the accelerating expansion of the universe. Dark energy was largely discovered through the 2011 Nobel Prize-winning work of Saul Perlmutter, an astrophysicist still active at Berkeley Lab. The name of this supercomputer system is named after this scientist.

Perlmutter’s Stage I cabinets do not have an attached door, revealing the blue and red lines of the direct liquid cooling system.

Projects with a similar spirit will be launched on this new supercomputer. For example, jobs in materials science aim to explore atomic interactions in order to find a way to create better batteries and biofuels.

Traditional supercomputers can barely handle the math needed to create simulations of a few atoms in nanoseconds with programs like Quantum Espresso. But by combining their highly accurate simulations with machine learning systems, scientists can study more atoms over longer periods of time.

“In the past it was not possible to do fully atomic simulations of large systems like the battery interface, but now the scientists plan to use Perlmutter to do it,” said Brandon Cook, a application implementation experts at NERSC share.

At the online launch event, NVIDIA CEO Jensen Huang congratulated on his plans to advance science with supercomputers.

Perlmutter’s ability to combine AI and high-performance computing will lead to breakthroughs in areas ranging from materials science and quantum physics to climate forecasting, biological research and more, he said. again.”

Researchers who want to do their work on Perlmutter can now submit access requests to the system.