US Takes Fastest Supercomputer Title With IBM's Summit

The competition with China continues to grow as the country widened its lead in total number of systems and Lenovo became the top vendor on the list.

IBM Summit

The latest edition of the biannual Top500 list of the world’s fastest supercomputers put into stark focus the increasingly heated competition between the United States and China that could set the direction of the high-performance and enterprise computing markets down the road, particularly as the world marches toward exascale computing.

According to the list, which was released this week at the International Supercomputing (ISC) 2018 show in Frankfurt, Germany, the Summit supercomputer (pictured) built by IBM and housed at the Oak Ridge National Laboratory is now the world’s fastest, wresting the title from China’s Sunway TaihuLight, which had sat atop the list for the past two years. The 4,356-node Summit—with each node powered by two 22-core Power9 chips and six Nvidia Tesla V100 GPU accelerators and linked via a Mellanox EDR InfiniBand network—delivers a performance of 122.3 petaflops, more than the 93 petaflops of the Sunway TaihuLight.

Summit puts a United States system at the top of the list for the first time since 2012, when China grabbed the position and kept hold of it for about five years. In addition to Summit, the introduction of Sierra—another IBM-built supercomputer for a national lab, this one the Lawrence Livermore Lab—into the No. 3 position was another win for the country. Sierra is armed with similar components as Summit, including Power9 chips, Nvidia V100 GPUs and Mellanox EDR InfiniBand interconnect.

Overall, the United States claimed six of the top 10 spots on the list and, according to the Top500 organizers, all of the U.S.-based systems account for 38.2 percent of the aggregate installed performance on the list. That position used to be held by Chinese supercomputers, which now are in second with 29.1 percent.

All that said, China continued to flex its growing computing muscle. The number of total systems on the list based in China grew from 202 in November 2017 to 206, while the number of U.S. systems fell from 145 to 124. At the same time, Lenovo—which is based in China but also has corporate offices in North Carolina—overtook Hewlett Packard Enterprise with the most installed systems on the list at 122, or 23.8 percent.

HPE was next at 15.8 percent (79 systems), followed by Inspur (of China) with 13.6 percent, Cray (United States) at 11.2 percent and Sugon (China) with 11 percent. Only the three companies from China saw percentage gains over the last list.

Made in China

The competition between the two countries has ramped up in recent years. Chinese officials have made no secret of their plans to make the country a technology powerhouse by investing billions of dollars to drive China-based companies to develop products for use in Chinese systems. The country’s Made in China 2025 campaign, which was announced in 2015, targets 10 high-tech sectors, including semiconductors, artificial intelligence (AI) and robotics, in which the country wants to lead.

The Sunway TaihuLight, which is housed in the National Supercomputing Center in Wuxi, is made up of a number of Chinese-made components, including the Sunway SW26010 processors, Sunway interconnect and Sunway operating system.

The program also is part of the growing trade dispute between the two companies, with some products related to the Made in China 2025 effort part of the United States’ $50 billion in tariffs on Chinese goods. Chinese officials reportedly are now trying to downplay the Made in China 2025 initiative in hopes of alleviating tensions with the Trump administration, which now is threatening another $450 billion in tariffs.

Exascale Computing

The rush to exascale computing is another area of the U.S.-China competition. China has three pre-exascale systems that are being prepared that will be based on three different architectures, much of them based on home-grown technologies. Meanwhile, the United States plans to roll out the first of three exascale systems in 2021.

Having a leading position in exascale computing will be important for whatever country gets there. The systems hold the promise of 50 times the speed of the most powerful supercomputers today, and challenges like increasingly complex applications such as data analytics, AI and machine learning and the slowing of Moore’s Law are driving demand for exascale systems. The country with the edge in exascale computing will have an advantage in everything from scientific and military R&D, national defense and business innovation.

Other highlights from the Top500 list include the continued growth in the number of systems that use GPU accelerators—mostly from Nvidia—to drive performance and power efficiency in the supercomputers and to handle the increase in workloads that demand parallel processing. Accelerators are used in 110 of the Top500 systems—up from 101 in November—with 96 of them and five of the top 10 running Nvidia GPUs. Seven use Intel’s Xeon Phi x86 coprocessors and four use Pezy accelerators.

In a post on the Nvidia blog, Ian Buck, vice president and general manager of the vendor’s Tesla Data Center business unit, noted that most of the new processing power on the Top500 list comes from GPUs, which also provide 95 percent of the flops of the Summit supercomputer. The system holds 27,648 Volta Tensor Core GPUs.

“As Moore’s Law continues to slow, accelerated computing clearly emerged at ISC as the booster rocket that will soon propel us into the age of exascale computing,” Buck wrote. “Every country is racing to build exascale systems. Peek at the Top500 list of 2025 and you’ll likely see over a dozen of such systems, with multi-precision accelerated computing the platform of choice. By comparison, all the systems added together on this week’s new Top500 list barely achieved an exaflop of total computing. This speaks to the massive opportunity ahead.”

As Moore’s Law slows down, performance gains will have to come from somewhere other than simply the CPU, he wrote, adding that GPUs will help drive those gains.