Amidst a flurry of activity surrounding machine learning, quantum and cloud, Tuesday’s Hyperion Research briefing at ISC 2018 reminded us that there is no escaping the race to exascale. And while Hyperion’s estimation that China will beat out the United States, Japan and Europe comes as no surprise to those keeping tabs, CEO Earl Joseph was quick to point out that the exascale panorama is in almost constant flux.
Most recently that landscape was disrupted by the U.S. announcement of its intention to procure two or potentially three exascale systems under the CORAL-2 program, with an anticipated cost of up to $600 million per supercomputer. That’s in addition to the retooled Intel-Cray Aurora system that is projected to reach over one exaflops peak. And between all four of the global players, Hyperion reports that they’re seeing shake ups every four to six weeks.
But as it currently stands, Joseph says that we should expect to see China reach the first peak and sustained exascale systems, with peak estimated to arrive in 2020, and sustained in either 2021 or 2022, with the U.S. expected to trail China by 6-9 months, crossing the same finish lines in 2021 and 2022-2023, respectively.
China’s proposed systems are expected to source hardware and processors from Chinese vendors (with a possibility of some U.S. processors included in the mix). The potential systems in contention are: Sugon Exascale, Sunway Exascale, TianHe-3, and potentially a Wuxi system.
Meanwhile, Hyperion lists four target systems in the U.S. lineup: ANL’s A21, ORNL’s Frontier (OLCF5), LLNL’s El Capitan (ATS-4), NERSC-10. System deliveries are expected to begin in 2021 and will have roughly one year between installations, with early operation expected one year later for each system. Hyperion also included a placeholder for an NSF Exascale Phase 2 system, which is slated to achieve a 10x performance boost over the Phase 1 machines. The systems are expected to feature American hardware and processors, with the potential inclusion of Arm.
But Europe and Japan have been a bit trickier to pin down, as Hyperion COO Steve Conway stepped in to explain, particularly as it related to Europe. With a growing emphasis on indigenous processors across the board, the European exascale effort is at a significant disadvantage. In addition to contending with red tape, selecting the host country, and budget concerns, a European plan to design and develop its own processors and accelerators has changed the schedule significantly, and is likely to cause the EU exascale timeline to slip by 2-4 years, according to the Hyperion analysis.
Conway noted that, to date, approximately 20 EU member states have committed to help fund the €1 billion effort. ETP4HPC and the CEA-RIKEN collaboration around ARMv8 are expected to be major contributors to the effort.
By comparison, Conway remarked that Japan was the most stable of the four contestants. Backed by a well-established processor and facing budgetary concerns, Hyperion expects that rather than racing to be the first, Japan is likely looking to climb to the number-one spot after the race has ended by leveraging a system with a traditional architecture and extreme bandwidth.
Looking at the numbers, investments in R&D among the group bore out Hyperion’s estimates, with the U.S. commitments at $2 billion per year, China in the same ballpark, Europe planning investment of €5-6 billion in total, and Japan potentially investing $1 billion over five years. But if you look at historical investments in HPC, it’s worth noting that European investments have gained significant momentum, jumping from a 25 percent share of global HPC spending to 29 percent, and Conway added that funding is increasing dramatically across the board.
*Data compiled through a combination of publicly available data and Hyperion’s estimates.