A ribbon-cutting ceremony held virtually at Berkeley Lab’s National Energy Research Scientific Computing Center (NERSC) today marked the official launch of Perlmutter – aka NERSC-9 – the GPU-accelerated supercomputer built by HPE in partnership with Nvidia and AMD. The HPE Cray EX supercomputer harnesses 6,159 Nvidia A100 GPUs and ~1,500 AMD Milan CPUs to deliver nearly 3.8 exaflops of theoretical “AI performance” (see endnote) or about 60 petaflops of peak double-precision (standard FP64) HPC performance.
The system is the namesake of Saul Perlmutter, an astrophysicist at Berkeley Lab who shared the 2011 Nobel Prize in Physics for his contributions to research showing that the expansion of the universe is accelerating. So it’s fitting that one of the initial use cases for the Perlmutter supercomputer will be in support of the Dark Energy Spectroscopic Instrument (DESI), which is probing the effect of dark energy on the universe’s expansion.
The Perlmutter system will help map the visible universe spanning 11 billion light years by processing data from DESI, which is capable of capturing as many as 5,000 galaxies in a single exposure.
In order to know where to point this expensive instrument each evening, researchers need to assess the data from the night before. Perlmutter can analyze dozens of exposures quickly enough to provide this feedback in time for the next nightly cycle.
In early benchmarking, NERSC researchers have reported up to 20X performance speedups using the GPUs, which they say will accelerate their workflows from a matter of weeks or months down to hours.
Materials science is expected to see similar benefits, laying the way for advances in batteries and biofuels. Applications such as Quantum Espresso leverage Perlmutter’s traditional simulation and machine learning capabilities, enabling scientists to study more atoms over a longer time period.
“In the past it was impossible to do fully atomistic simulations of big systems like battery interfaces, but now scientists plan to use Perlmutter to do just that,” said Brandon Cook, an applications performance specialist at NERSC.
Nvidia reported that Quantum Espresso, BerkeleyGW and NWChem all are capable of leveraging Nvidia’s FP64 Tensor Cores, unlocking double the performance of the standard FP64 format — 19.5 teraflops versus 9.7 teraflops (peak theoretical) per GPU. (Nvidia reports that Perlmutter provides 120 petaflops of peak FP64 Tensor Core performance.)
The first phase of Perlmutter spans 12 GPU-accelerated Cray EX cabinets (aka “Shasta”) housing more than 1,500 nodes and 35 petabytes of all-flash parallel file system (HPE E1000). The Lustre filesystem will move data at a rate of more than 5 terabytes/sec making it the fastest storage system of its kind, according to NERSC.
The Perlmutter system is direct liquid cooled and uses HPE’s Cray-developed Slingshot interconnect technology.
A second CPU-only phase is planned for later this year. Phase 2 adds 12 CPU cabinets with more than 3,000 nodes, equipped with two AMD Milan CPUs with 512GB of memory per node. The Phase 2 system also adds 20 more login nodes and four large memory nodes, according to NERSC.
Perlmutter is the successor to Cori (named in honor of Nobel Prize-winning biochemist Gerty Cori), which was also constructed as two partitions, the Phase 1 Intel Haswell-based “Data Partition” and the Phase 2 Intel Knights Landing (Xeon Phi) partition. Cori is the largest supercomputing system for open science based on KNL processors. NERSC will continue to operate Cori through at least 2022.
On the software side, Perlmutter users will have access to the standard NVIDIA HPC SDK toolkit, and support for OpenMP is forthcoming through a joint development effort with NERSC.
Python programmers will be able to use RAPIDS, Nvidia’s open software suite for GPU-enabled data science.
Phase 1 cabinets were deployed over the last few months but even before installation began in November 2020, the NERSC Exascale Science Applications Program (NESAP) was engaged in readiness activities to be able to leverage the GPU nodes for simulation, data, and learning applications starting on day one. NERSC reports that these NESAP readiness teams will be the first to access the system. Support for Exascale Computing Project (ECP) software is also planned on the new system.
AI for Science
With its strong AI capabilities, Perlmutter ties into the DOE’s AI for Science focus area, an exascale-like initiative for advancing the use of AI in science.
“AI for science is a growth area at the U.S. Department of Energy, where proof of concepts are moving into production use cases in areas like particle physics, materials science and bioenergy,” said Wahid Bhimji, acting lead for NERSC’s data and analytics services group, in an Nvidia blog post.
“People are exploring larger and larger neural-network models and there’s a demand for access to more powerful resources, so Perlmutter with its A100 GPUs, all-flash file system and streaming data capabilities is well timed to meet this need for AI,” he added.
Presenting in a pre-recorded video as part of today’s virtual launch program, Nvidia CEO Jensen Huang underscored emerging HPC and AI synergies.
“Perlmutter’s ability to fuse AI and high performance computing will lead to breakthroughs in a broad range of fields from materials science and quantum physics to climate projections, biological research and more,” Huang said.
Looking Ahead (to Quantum)
Planning is already underway for the follow-ons to Perlmutter, codenamed NERSC-10 and NERSC-11.
“Systems take years and years for us to design and deploy,” said NERSC Director Sudip Dosanjh during today’s virtual dedication ceremony.
“It’s pretty clear that we’ll have more heterogeneous systems as we enter the post-Moore’s law era. We’re looking at different types of accelerators. I don’t think that it’s likely that NERSC-10 will have a quantum accelerator, but NERSC-11 certainly might. Half the codes that run at NERSC solve some kind of quantum mechanical problem, and that part of the workload might really benefit from a quantum accelerator.
“With NERSC-10, we’re really going to focus on end-to-end DOE Office of Science workflows, and hopefully enable new modes of scientific discovery through the integration of experiment, data analysis and simulation. And so not only do we want to make sure that the scientists can use AI to analyze the data, but we also want to use AI to manage the system to increase the reliability of the system and the energy efficiency of the system. And in addition we have a goal of using AI to reconfigure NERSC-10 to accelerate workflows,” said Dosanjh.
Hello Perlmutter — Saul Perlmutter inaugurates Perlmutter in a live demo:
Note: Perlmutter’s “AI performance” is based on Nvidia’s half-precision numerical format (FP16 Tensor Core) with Nvidia’s sparsity feature enabled.