Additional details came to light on Argonne National Laboratory’s preparation for the 2022 Aurora exascale-class supercomputer, during the HPC User Forum, held virtually this week on account of the pandemic.
Exascale Computing Project Director Doug Kothe reviewed some of the ‘early exascale hardware’ at Argonne, Oak Ridge and NERSC (Perlmutter), while Ti Leggett, deputy project director & deputy director of operations at Argonne National Laboratory, gave an account of the Polaris system that was announced last month and is currently being installed at the lab.
Arcticus is the name of the early access machine at the Argonne Leadership Computing Facility (ALCF) with the Intel Xe-HP GPU nodes that was delivered to the lab last fall. The 17-node system is serving as a transitional development vehicle for the future (2022) Aurora supercomputer, subbing in for the forthcoming Intel Xe-HPC (“Ponte Vecchio”) GPUs that are the computational backbone of the system. Other early hardware at ALFC includes systems Iris and Yarrow, as well as a single rack Cray EX testbed, called Crux.
Meanwhile, at Oak Ridge Leadership Computing Facility (OLCF), Spock is helping researchers prepare for the heterogeneous HPE-AMD Frontier exascale supercomputer, arriving this year. Spock is an HPE Apollo 6500 Gen10 Plus system with 12 nodes, each outfitted with one 64-core AMD Epyc Rome CPU and four MI100 GPUs.
“All of our teams, both apps and software, are getting on, getting experiences,” said Kothe. “Not surprisingly, [they are] finding issues, features we need, or bugs we need to fix. There’s been a lot of interaction with our our vendors, in this case, Intel, HPE and AMD.”
Offering further perspective into the recently announced Polaris supercomputer (44 double-precision petaflops peak), Leggett said the HPE Apollo system would provide a bridge from Theta, Argonne’s current manycore Intel Knights Landing platform, to Aurora, the heterogeneous CPU-GPU based HPE XE platform.
As we covered previously, Polaris – currently under construction at Argonne – combines 560 AMD Epyc Rome CPUs and 2,240 Nvidia A100 GPUs into 40 HPE Apollo Gen10 racks, connected by HPE’s Slingshot networking.
Polaris will start out with HPE Slingshot 10 with a planned upgrade to Slingshot 11, when it becomes available next year. Notably Slingshot 11 increases performance from 100 Gbps to 200 Gbps, and is the same technology that Aurora will use. Slingshot 11 also provides greater MPI optimized collectives and optimizations, said Leggett.
Polaris’s Slingshot networking will be organized into 11 local Dragonfly groups, 10 of those will be compute and the 11th has the non-compute nodes, said Leggett. Each compute group contains four racks, and each rack has seven Apollo chassis with two nodes per chassis. In total, each local group has 56 nodes, connected together in the Dragonfly topology. The 11th group contains the management and gateway nodes, with the latter providing the connection to the Lustre file systems (Grand and Eagle).
Leggett concluded his brief talk with a summary of the ways that Polaris provides an on-ramp to Aurora, illustrated by this chart:
“Everything that’s in bold is directly related to technologies that will be used on Aurora,” said Leggett. “You can see there’s a quite a bit of overlap, and that’s good. Even though the multi-GPU environment isn’t specifically bold, it’s a very similar platform. And so I think it will be a very fitting resource for helping our users to scale up and be prepared for exascale in 2022.”
Aurora was further detailed at HPC User Forum by Intel chief architect and Aurora technical lead Robert Wisniewski who presented the slide below. (He noted that the number of racks is NOT an exact count.)
Aurora’s compute nodes will be powered by Intel’s Sapphire Rapids (“Intel 7” node) CPU and Xe Ponte Vecchio GPU. With Xe cores manufactured on TSMC’s N5 process technology, the Ponte Vecchio GPU A0 silicon provides ~45 theoretical teraflops of single- or double-precision performance (it was stated during the Hot Chips conference that the architecture rate is the same for FP32 and FP64).
Factoring a conservative performance efficiency of, say, 70 percent, a Ponte Vecchio GPU might provide 31.5 HPL teraflops each(!). Aurora system design specs, however, have per-node performance at >130 teraflops, which comes out to “only” about 21 peak teraflops per GPU. Multiplied by 9,000(+) nodes, that is 1.17(+) exaflops. Peak power for Aurora is listed as ≤ 60MW, more than twice the spec’d power footprint of Frontier.