Three major forces – AI, cloud and exascale – are combining to raise the HPC industry to heights exceeding expectations. According to market study results released this week by Hyperion Research at SC19 in Denver, HPC server sales in the first half of the year totaled $6.7 billion, while 2018 sales grew 15 percent overall, totaling $13.7 billion. Growth has been driven primarily by new buyers from the enterprise moving into HPC for AI-related workloads, such as fraud detection, business intelligence, affinity marketing, personalized medicine, smart cities and IoT, the firm said.
Hyperion expects the good times to keep rolling for HPC server vendors, with that industry sector projected to reach just under $20 billion by 2023 for a 2018-2023 CAGR of 7.8 percent. For the broader HPC market, encompassing servers, storage, middleware, applications and services, revenue is expected to reach slightly less than $40 billion for a CAGR of 7.2 percent.
Other key takeaways from Hyperion’s industry update:
- HPE maintained its significant lead in HPC server sales, not including revenues from Cray, acquired by HPE in May, with $4.8 billion in revenue for a 34.8 percent share, ahead of Dell EMC’s share of 20.8 percent, followed by IBM, Lenovo and Inspur.
- View by vertical markets, government laboratory applications was the largest sector in 2018 with $2.6 billion in server sales, followed by university/academic at $2.4 billion, defense at $1.4 billion and computer-aided engineering at $1.5 billion.
- High Performance Data Analytics (HPDA)-AI is growing faster than the overall HPC market, and the AI subset is growing faster than HPDA, though in absolute dollars it will remain smaller. Hyperion reported that worldwide HPC-based AI (machine/deep learning and other) is expected to reach $2.7 billion by 2023 for a 2018-2023 CAGR of 29.5 percent while worldwide HPDA revenues are expected to reach $6.45 billion by 2023, a five-year CAGR of 15.4 percent.
Yet for all that, Hyperion Senior Research VP Steve Conway said AI is just starting.
“There’s pretty good consensus that we are at a period that we’re still near the beginning with AI as far as capability goes,” he said. “Lots of emphasis has been on training and there’s been lots of progress there. Just more recently, the emphasis is starting to shift to inferencing, the ability of the machines to guess better and to make better intelligence, intuition. But that’s still not so far along.”
- Hyperion declared 2019 a tipping point year for a significant and long-anticipated shift in market attitudes toward running HPC workloads in clouds, resulting in Hyperion increasing its revenue forecast from $3 billion to $4 billion for this year and totaling $7.5 billion by 2023. By applications, HPC in the cloud will be led by bio-sciences, with a 2018-2023 CAGR of 25.6 percent, followed by CAE at 24.7 percent and chemical engineering at 23.9 percent.
Based on recent surveys of HPC cloud users reporting that they run 33 percent of their HPC workloads in 3rd-party clouds, Hyperion said the HPC community runs 20 percent of workloads in cloud environments.
“That is a very big jump from 18 to 24 months ago, that number started at just under 10 percent,” Conway said. Having said that, however, he cautioned that HPC in clouds has limitations, including moving mission critical workloads off-prem and high costs associated with data locality where large volumes of data are involved.
- Two of the first three American exascale (a billion billion calculations per second) systems are expected to be delivered by late next year, with acceptance in 2022. These are Aurora, comprised of the Cray Shasta architecture with Intel Xeons and Intel Xe GPUs (see our coverage of Intel’s SC19 disclosures), for Argonne National Laboratory; and Frontier, comprised of Cray Shasta and AMD Epyc CPUs and future Radeon GPUs for Oak Ridge National Lab. The third system, El Capitan, based on the Cray Shasta architecture, is expected to be delivered to Lawrence Livermore National Lab within two years, with acceptance in 2023. Overall, major government-sponsored efforts will drive development of about 26 near-exascale and exascale systems by 2025, with total spending of about $9 billion, Hyperion said. After that, exascale activity may decline sharply.
“It is unclear to what extent future (beyond 2025) large supercomputers will continue to be purchased in the $500 million-plus price range,” the firm said, “or will return to more ‘modest’ price levels in the $100 to $250 million price range per system.”
As for the international exascale race, Hyperion sees the U.S. and China installing the first exascale systems in similar timeframes – by 2022 – followed by the EU and then Japan.
Hyperion also announced winners of its annual Innovation Excellence Awards, including:
- Argonne National Laboratory used high spatial resolution physical-based models for regional climate, surface hydrology and coastal flooding to estimate flooding risks, used by AT&T to develop a risk and resiliency tool.
- The Max Delbruck Center for Molecular Medicine worked with the Zuse Institute, Berlin, to combine convolutional neural networks and 3D statistical shape knowledge on a DDN A3I system to automate segmentation of knee bone and cartilage for identification of osteoarthritis.
- For innovative AI hardware, Hyperion recognized Andrew Feldman of Cerebras and its Wafer Scale Engine containing more than 1.2 trillion transistors and 400,000 compute cores “to solve AI problems, that used to be solved in months, in minutes.”
- For innovative HPC system design, Hyperion recognized Sugon for its development of a silicon cube blade system using an efficient, phase-change liquid cooling technology for a system with high computing density and energy efficiency.