Hyperion Research – formerly IDC’s HPC group – yesterday painted a fascinating and complicated portrait of the HPC community’s health and prospects at the HPC User Forum held in Albuquerque, NM. HPC sales are up and growing ($22 billion, all HPC segments, 2016). Global exascale plans are solidifying (who, what, when, and how much ($)). The new kid on the block – all things ‘big’ data driven – is becoming an adolescent and behaving accordingly. And HPC ROI, at least as measured by Hyperion, is $551 per $1 invested (revenue growth) and $52 per $1 of profit invested.

This new version of HPC has been taking shape for some time and most of the themes are familiar (see HPCwire 2015 article, IDC: The Changing Face of HPC): industry consolidation, SGI’s acquisition by HPE along with the Dell EMC merger being the most recent; accelerated computing versus Moore’s Law; the growing appetite of HPC technology suppliers for expansion into the enterprise; big data’s transformation into a more nuanced multi-faceted blend of technologies and applications making it a form of HPC. These are just a few of the major trends laid out by Hyperion at its HPC User Forum.

All netted down, HPC is still expected to be a growth market, according to Earl Joseph, now CEO of Hyperion, which is expected to be acquired by year’s end. Joseph cited the following drivers:

• Growing recognition of HPC’s strategic value.
• HPDA, including ML/DL, cognitive and AI.
• HPC in the cloud will lift the sector writ large.

“There’s a lot of growth in the upper half of the market and we are back to slowdown in the lower half of the market,” said Joseph. “Supercomputers are showing a very good recovery but they still haven’t hit the high point (~$5 billion) of three or four years ago.” They likely won’t get back to that level till 2022/2023 suggested Joseph.

Overall the HPC market segments have tended to hold their position. Storage ($4,316 million) remained the largest non-server segment and the fastest growing segment overall with a 7.8 percent annual growth expected over the next five years.

Vendor jockeying will continue he noted. Consolidation has been a major factor. HPE topped the revenue list in 2016 and will likely do so again in 2017 when SGI’s revenue is added. Dell EMC would no doubt question that and it will be interesting to watch this rivalry. IBM has never recovered its position after jettisoning its x86 businesses. The battle between x86 offerings, IBM Power, and ARM continues with both Europe and Japan making substantial bets on ARM for HPC uses. Indeed, the rise of heterogeneous computing generally is creating new opportunities for a variety of accelerators and accelerated systems.

These are the top HPC server suppliers by revenue ($ millions) according to Hyperion: HPE/HP ($3,878), Dell ($2,014), Lenovo ($909), IBM ($492), Cray ($461), Sugon ($315), Fujitsu ($226), SGI ($169), NEC ($166), Bull Atos ($118), and Other ($2,453). Interesting to note that “Other” is the second largest total revenue.

Not surprisingly, Hyperion looked closely at the intensifying race for exascale machines. China, for example, has three efforts on the path to exascale. Joseph expects China to be first to stand up an exascale. “They are saying 2019 but we’re not sure they will hit that date. We’re saying 2020,” said Joseph. The major players – U.S., EU, Japan, and China – are all speeding up their efforts. In the U.S., for example, Path Forward awards are expected soon.

Many questions remain. China is still selecting final vendors, something that was supposed to be done last fall said Joseph. Japan’s design is the closest to being “locked in” with the prime contractor Fujitsu having settled on an ARM-based architecture. But that project has experienced some delay and its financing method is not fixed.

“According to Japan’s latest announcement, their machine will be up in 2023 but we really expect it to be 2024. The cost may be a bit higher too, $800 million to $900-plus million range. Also, the Japanese government has not yet agreed to fund the whole system. They are funding it one year at time,” said Joseph.

Nevertheless, exascale funds are starting to flow and plans are taking firmer shape. As shown here, Hyperion has characterized the major exascale programs and forecast likely costs, technology choices, and timetables. Paul Messina, director of the U.S. Exascale Computing Project, provided an update at the HPC User Forum and HPCwire will have detailed coverage of the U.S. effort shortly.

Predictably, the Hyperion presentation covered a lot of ground drawn from Hyperion/IDC’s ongoing research efforts. Steve Conway, another IDC veteran and now Hyperion SVP research, reviewed the adoption of HPDA as well as zeroing in on two of its drivers, deep learning and machine learning. You may recall that IDC was one of the first to recognize the rise of data analytics as part of HPC. Clearly there are many potential uses cases Conway said. Today, the HPC-HPDA convergence is taken for granted and is depicted in the slide below.

Hyperion has just created four new data-intensive segments, bulleted here, with more to follow:

• Fraud and anomaly detection. Two example use cases include government (intelligence, cyber security) and industry (credit card fraud, cyber security).
• Affinity Marketing. Discern potential customers’ demographics, buying preferences and habits.
• Business intelligence. Identify opportunities to advance market position and competitiveness.
• Precision Medicine. Personalized approach to improve outcomes, control costs.

“Fraud and anomaly detection are the largest today. Business intelligence is growing quickly. The tortoise that will probably win the race is precision medicine because of the size of the health care over time,” said Conway, noting the HPDA market is growing two to three times faster than traditional overall HPC market.

Not surprisingly, deep learning is the darling of this frontier and also the most technically challenging. Singling out precision medicine as a promising area for DL, Conway said “IBM Watson is the name that’s known here but I promise you x86 clusters are doing the same thing.”

Making the machine learning to deep learning shift is a difficult journey said Conway. Having enough data both to train deep learning systems and also to infer high fidelity decisions when put into practice is the big challenge. “If you are in the realm of Google or Baidu or Facebook, you have plenty of data. If you are outside of that realm you are in trouble. In most of these realms you do not have enough data to do deep learning,” said Conway.

“One case in point, and we have many of them: We talked to the United Health Group which has about 100 million people that it covers; that’s not nearly enough to do the deep learning they need and they know it. They have built a facility in Cambridge, Mass., and invited competitors to come in and to pool anonymized data to try to get to the point where they can actually start playing with deep learning. This is a big issue.”

Aside from having enough data, there’s the computation challenge. Today, GPUs “rule the roost in these ecosystems, with the software built around them, but we expect to see other things like Intel Phis and the remarkable resurgence of FPGAs have a role. Another big issue vendors are having here is there really aren’t good benchmarks and they spend too much time just trying to decide what would be satisfactory results,” Conway said.

In earlier studies HPC user willingness to deploy in the cloud has often seemed tepid. Costs, security, adequate performance (data movement, computation, and storage) were all concerns, especially so in public cloud. Hyperion suggested attitudes seem to be changing and reported a jump in the number of HPC sites using public clouds – 64 percent now up from 13 percent in 2011. Conway cautioned that the size and number of jobs were still limited to a small proportion of any give user’s needs. Conversely, suggested Conway, private and hybrid cloud use was growing fast and held more near-term promise.

Despite the great flux within HPC many areas have changed little according to Hyperion. For example, software problems (management s/w, parallel s/w, license issues, etc.) remain the number one pain point to HPC adoption or use according to Hyperion research. This prompted a member of the audience to say, “Earl, this looks like exactly the same IDC slide I saw ten years ago.” It sort of is.

Storage access time was now the number two complaint, followed by clusters still too hard to use and manage.

Hyperion presented a fair amount of detail concerning its ROI study and is making the full data available to requesters. (Download Results: www.hpcuserforum.com/ROI)

Slides courtesy of Hyperion Research.