The Frontier supercomputer on the Division of Power’s Oak Ridge Nationwide Laboratory has earned the highest rating because the world’s quickest on the 59th TOP500 checklist, with 1.1 exaflops of efficiency. The system is the primary to realize an unprecedented degree of computing efficiency referred to as exascale, a threshold of a quintillion calculations per second.
The Frontier supercomputer’s exascale efficiency is enabled by a number of the world’s most superior items of know-how from HPE and AMD:
-
Frontier has 74 HPE Cray EX supercomputer cupboards, that are purpose-built to help next-generation supercomputing efficiency and scale, as soon as open for early science entry.
-
Every node incorporates one optimized AMD EPYC processor and 4 AMD Intuition accelerators, for a complete of greater than 9,400 CPUs and greater than 37,000 GPUs in your complete system. These nodes present builders with simpler capabilities to program their purposes, because of the coherency enabled by the EPYC processors and Intuition accelerators.
-
HPE Slingshot, the world’s solely high-performance Ethernet material designed for next-generation HPC and AI options, together with bigger, data-intensive workloads, to handle calls for for greater velocity and congestion management for purposes to run easily and enhance efficiency.
-
An I/O subsystem from HPE that may come on-line this 12 months to help Frontier and the OLCF. The I/O subsystem options an in-system storage layer and Orion, a Lustre-based enhanced center-wide file system that can also be the world’s largest and quickest single parallel file system, based mostly on the Cray ClusterStor E1000 storage system. The in-system storage layer will make use of compute-node native storage units linked by way of PCIe Gen4 hyperlinks to offer peak learn speeds of greater than 75 terabytes per second, peak write speeds of greater than 35 terabytes per second, and greater than 15 billion random-read enter/output operations per second. The Orion center-wide file system will present round 700 petabytes of storage capability and peak write speeds of 5 terabytes per second.
-
As a next-generation supercomputing system and the world’s quickest for open science, Frontier can also be energy-efficient, because of its liquid-cooled capabilities. This cooling system promotes a quieter datacenter by eradicating the necessity for a noisier, air-cooled system.
Frontier includes a theoretical peak efficiency of two exaflops, or two quintillion calculations per second, making it ten instances extra highly effective than ORNL’s Summit system. The system leverages ORNL’s in depth experience in accelerated computing and can allow scientists to develop critically wanted applied sciences for the nation’s power, financial and nationwide safety, serving to researchers deal with issues of nationwide significance that had been inconceivable to resolve simply 5 years in the past.
Rankings had been introduced on the Worldwide Supercomputing Convention 2022 in Hamburg, Germany. Frontier’s speeds surpassed these of every other supercomputer on this planet, together with ORNL’s Summit, which can also be housed at ORNL’s Oak Ridge Management Computing Facility, a DOE Workplace of Science consumer facility.
Frontier, a HPE Cray EX supercomputer, additionally claimed the primary spot on the Green500 checklist, which charges power use and effectivity by commercially out there supercomputing techniques, with 62.68 gigaflops efficiency per watt. Frontier rounded out the twice-yearly rankings with the highest spot in a more moderen class, mixed-precision computing, that charges efficiency in codecs generally used for synthetic intelligence, with a efficiency of 6.88 exaflops.
The work to ship, set up and check Frontier started through the COVID-19 pandemic, as shutdowns all over the world strained worldwide provide chains. Greater than 100 members of a public-private staff labored across the clock, from sourcing hundreds of thousands of parts to making sure deliveries of system elements on deadline to rigorously putting in and testing 74 HPE Cray EX supercomputer cupboards, which embody greater than 9,400 AMD-powered nodes and 90 miles of networking cables.
Frontier’s general efficiency of 1.1 exaflops interprets to a couple of quintillion floating level operations per second, or flops, as measured by the Excessive-Efficiency Linpack Benchmark check. Every flop represents a doable calculation, corresponding to addition, subtraction, multiplication or division.
Frontier’s early efficiency on the Linpack benchmark quantities to greater than seven instances that of Summit at 148.6 petaflops. Summit continues as a formidable, extremely ranked workhorse machine for open science, listed at quantity 4 on the TOP500.
Frontier’s mixed-precision computing efficiency clocked in at roughly 6.88 exaflops, or greater than 6.8 quintillion flops per second, as measured by the Excessive-Efficiency Linpack-Accelerator Introspection, or HPL-AI, check. The HPL-AI check measures calculation speeds within the computing codecs sometimes utilized by the machine-learning strategies that drive advances in synthetic intelligence.
Detailed simulations relied on by conventional HPC customers to mannequin such phenomena as most cancers cells, supernovas, the coronavirus or the atomic construction of components require 64-bit precision, a computationally demanding type of computing accuracy. Machine-learning algorithms sometimes require a lot much less precision—typically as little as 32-, 24- or 16-bit accuracy—and might reap the benefits of particular {hardware} within the graphic processing items, or GPUs, relied on by machines like Frontier to achieve even sooner speeds.
ORNL and its companions proceed to execute the bring-up of Frontier on schedule. Subsequent steps embody continued testing and validation of the system, which stays on observe for ultimate acceptance and early science entry later in 2022 and open for full science initially of 2023.
