By reaching over 1.102 ExaFlop/s on a continuous Linpack run, the AMD-powered Frontier supercomputer has become the first exascale supercomputer to be formally recognized by the international community. With a huge increase in the number of AMD-powered systems on the list this year, it tops the newly-released Top500 list of the world’s fastest supercomputers. When compared to the previous leader, Japan’s Fugaku, Frontier not only surpasses it but also destroys it; in fact, Frontier is faster than the following seven supercomputers on the list, added together. Specifically, the system produces up to 1.69 ExaFlops in peak performance, with room to exceed 2 ExaFlops following further tuning, whereas Frontier reached 1.1 ExaFlops during a sustained Linpack FP64 benchmark. If you’re unfamiliar with the term, an ExaFlop is the rate at which one quintillion floating point operations may be performed in a second.
ExaFlops on the HPL-AI benchmark Performance
With a mixed-precision performance of 6.88 ExaFlops on the HPL-AI benchmark, Frontier is also the world’s fastest artificial intelligence system. That’s 68 million commands every second, or one for every one of the brain’s 86 billion neurons, demonstrating its formidable processing power. This machine appears to be in direct competition with Nvidia’s recently announced AI-focused supercomputers that are powered by Grace CPU Superchips built on the Arm architecture.
Frontier’s Design
Frontier’s design is now the most power-efficient supercomputing architecture in the world, as evidenced by the fact that the Frontier Test and Development (Crusher) system ranked first on the Green500 (the primary Frontier system ranks second on the Top500). During the qualifying benchmark run, the entire system delivered 52.23 GFlops per watt while only using 21.1 MW (megawatts). Frontier uses 29 Megawatts at full capacity.
Frontier supercomputer
Scale-wise, the Frontier supercomputer is impressive, but it’s only one of AMD’s numerous Top500 successes this year. As of this year’s ranking, five of the world’s top 10 supercomputers are powered by AMD EPYC systems, and ten of the top twenty. As of March 2022, 94 of the world’s Top500 supercomputers are powered by AMD’s EPYC, up from 73 in November 2021 and 49 in June of the same year. Also, this year more than half of the new systems on the list use AMD. As you can see in the accompanying photo gallery, most Top500 machines are still powered by Intel processors, while Nvidia graphics processing units remain the most popular accelerator.
Green500
The current Green500 list shows that Intel is still the leader in terms of overall performance, but AMD is clearly the leader when it comes to power efficiency, powering four of the world’s most energy-efficient computers and holding eight of the top ten and seventeen of the top twenty slots.
Oak Ridge National Laboratory
Oak Ridge National Laboratory (ORNL) in Tennessee is home to HPE’s Frontier supercomputer, which was built for the Department of Energy (DOE). Each of the 9,408 compute nodes in the system is outfitted with a 64-core AMD “Trento” processor, 512 GB of DDR4 memory, and four AMD Radeon Instinct MI250X graphics processing units. These nodes are housed in a total of 74 HPE Cray EX cabinets, each of which weighs in at a hefty 8,000 pounds on its own. The total amount of DDR4 memory in the system is 4.6 petabytes, and there are 602,112 CPU cores.
Ethernet-based Networking Fabric
There are also 8,138,240 cores and 4.6 petabytes of HBM memory spread over the 37,888 AMD MI250X GPUs (128GB per GPU). Using an Ethernet-based networking fabric, the central processing units and graphics processing units (GPUs) in the HPE Cray Slingshot-11 supercomputer are connected. The entire system employs direct water-cooling to keep things cold, with 350-horsepower pumps circulating 6,000 gallons of water (enough to fill an Olympic-sized pool in 30 minutes) through the system. Because the system doesn’t utilize chillers to cool the water, the water stays at a comfortable 85 degrees all the time.
Storage Subsystem
The storage subsystem is incredibly fast, with 15 billion IOPS, 700 petabytes of capacity, and 75 terabytes per second in throughput. The primary fast storage tier has 11.5PB of capacity and is made up of 5,400 NVMe SSDs, while the metadata tier has 10PB of space divided across 480 NVMe SSDs. At the same time, the 679PB of storage space is provided by 47,700 PMR hard drives.
ORNL faced its own unique difficulty in assembling Frontier due to the staggering quantity of components required (685 different part numbers and 60 million individual components). ORNL was short two million components after a chip shortfall hit during construction and affected 167 of those part numbers. Similar problems were experienced by AMD, which saw shortages in 15 MI200 GPU product numbers. To get around the shortages, ORNL collaborated with the ASCR to get the parts DPAS ratings, which means the US government used the Defense Act to acquire them because of the importance of Frontier to national defense.
Frontier’s Mechanical Plant
Frontier’s mechanical plant can cool up to 40 MW of computing power, which is the equivalent of 30,000 American houses, even though the system currently peaks at 29 MW of electricity. The facility has opportunity for expansion, since its capacity may be increased to 70 MW.
Although Frontier is regarded as the world’s first Exascale supercomputer, China is widely believed to be home to two Exascale supercomputers in the form of Tianhe-3 and OceanLight, both of which achieved the feat a year ago. In light of the current political climate between the United States and China, such systems have not been submitted to the Top500 committee. Some have questioned whether or not they are genuine exascale systems, at least as measured by an FP64 workload, due to the lack of formal submissions to the Top500; instead, a Gordon Bell submission was presented as a proxy.
World’s Fastest Supercomputer
For the time being, Frontier holds the title of world’s fastest supercomputer and is the first to achieve exascale. Aurora, powered by Intel and plagued by delays, is projected to go live later this year or early next year, challenging Frontier for the top spot in the supercomputing rankings with up to 2 ExaFlops of speed.
AMD’s next move? The El Capitan machine, which is expected to be a 2+ ExaFlop machine, is not expected to be operational until 2023 at the earliest. When finished, this supercomputer powered by Zen 4 will challenge Intel’s Aurora for the top spot in the Top500.
Meet Khurram Raheel Akbar our senior content writer. With over 10 years of experience in the field of content writing, Raheel, has established himself as an expert in creating engaging and informative content. His exceptional writing skills have enabled him to craft compelling stories that resonate with audiences across a variety of industries. Raheel’s writing style is concise, clear, and impactful, making him a go-to writer for any business looking to enhance its brand’s online presence. His dedication to staying up-to-date with the latest trends and strategies in content marketing allows him to create relevant and informative content that drives traffic and increases conversion rates. Raheel’s passion for writing is matched only by his commitment to delivering exceptional results to his clients. Whether you’re looking to revamp your website, launch a new product, or establish your brand’s voice, Raheel is the senior content writer you need to bring your vision to life.
