Researchers tapped GPUs in the world’s most powerful supercomputer to set a record simulating current flows through a chip, snagging the 2019 Gordon Bell prize, awarded at this week’s SC19 supercomputing show.
Their work could shed light on a wide variety of challenges in materials science, such as how to keep processors cool inside high-performance computers.
A team from ETH Zurich simulated how electricity flows through a 10,000-atom slice of transistors with similar accuracy and 14x faster than prior work on a 1,000-atom sample. At the heart of the work, researchers crafted a quantum transport solver that removed scaling bottlenecks, delivering results two orders of magnitude faster than previous efforts.
Nevertheless, the research required a whopping 90.89 petaflops of mixed-precision performance on Summit, the world’s top supercomputer, with two IBM Power 9 CPUs and six NVIDIA V100 Tensor Core GPUs per node. The work ran on 4,560 of Summit’s nodes, generating nearly half the system’s peak capability.
The simulation effectively created a map of where and how heat is generated and released in a single transistor. The work uses a data-centric approach to application development, which includes a graphical interface that lets users see via a visualization where data movement bottlenecks exist. Researchers hope it can lead engineers to new ways to design cooler chips and systems.
Today’s leading-edge processors pack billions of transistors in a few square millimeters of silicon, pulling hundreds of watts of power and the heat that comes with it. For example, when running full out, the Summit supercomputer consumes 13 megawatts.
The six-person team won the 32nd annual Gordon Bell Prize, given by the Association for Computing Machinery for outstanding work in high performance computing. Winners include Guillermo Indalecio Fernández, Torsten Hoefler, Mathieu Luisier, Tal Ben-Nun, Timo Schneider and Alexandros Nikolaos Ziogas.
Runners up for the prize awarded at SC19 were from the University of Michigan and also used Summit. Their work demanded 46 petaflops on the system and showed 200x speed improvements in density functional theory.