The HPC User Forum meeting taking place virtually this week (May 11-13) kicked off with Hyperion Research’s market update, covering the 2020 period. Although the HPC server market had been facing a 6.7 percent COVID-related dip in 2020, the inclusion of Fugaku system revenue in the fourth quarter brought the total HPC spend for the year to $13.7 billion, an increase of 1.1 percent over the previous year. On the horizon are continued growth for HPC cloud and a possible surge in storage spending.
Riken’s Fugaku supercomputer – the world’s fastest at 442 Linpack petaflops – was put into service a year ahead of schedule to fight COVID. Although the system did not reach full operational status until 2021, system revenue was recorded by Fujitsu in December 2020.
The shift in server market revenue from 2021 to 2020 portends a flat growth rate for for the next couple of years, said Hyperion Research CEO Earl Joseph. But the market will return to sizable growth, reaching roughly $19 billion by 2024, he said.
That puts the five-year, compound annual growth rate (2019-2024) for on-premises purchases at 6.8 percent, subject of course to COVID dynamics in addition to the usual market uncertainties. Because of COVID-19, Hyperion is adjusting its forecasts more regularly, making updates once or even twice a quarter, according to Joseph.
Looking at the market by system segment (supercomputer, divisional, departmental and workgroup), the supercomputer segment reached nearly $6 billion (up 13.7 percent from 2019). This robust growth was largely due to the the Fugaku system, which accounted for about $1 billion. The workgroup segment declined the most and is under pressure in multiple different ways, said Joseph.
By vertical, five segments hit or exceeded the $1 billion per year mark: government lab, university/academic, CAE, defense, and biosciences. Government lab was the highest of these, responsible for nearly $3.4 billion in HPC spending, with Fugaku contributing about a third of that.
Shifting to the vendor landscape, HPE and Dell Technologies maintained their first and second place positions, with 33.4 percent and 20.8 percent shares respectively, although server sales were down for both year-over-year. With the Fugaku sale on its books, Fujitsu had a strong year, earning $1.3 billion for a 9.6 percent share. Inspur is next with 7.2 percent market share, followed by Lenovo with 6.8 percent. The other category (which includes 39 companies) accounts for $1.5 billion in spending, a 10.9 percent share.
Adding in the rest of the on-prem market categories (storage, middleware, applications and service categories), Hyperion expects the total on-premises HPC market to grow to nearly $38 billion in 2024 (a CAGR of 6.6 percent).
For comparison, here is the market forecast that Hyperion issued last June, not-yet-revised for COVID effects. Note that the pandemic has seemingly caused or at least contributed to a two-year “growth lag” relative to the most recent pre-COVID forecast.
AI, Cloud, Storage & Exascale = Growth!
Certain subsegments of the market are growing faster than others, as much as seven times faster. The highest growth areas are related to AI, machine learning and deep learning, along with HPDA (high-performance data analytics) and big data, said Joseph. The use of deep learning in HPC is expected to grow the fastest with a 41.8 percent projected five-year CAGR. The traditional modeling and simulation area is expected to grow at roughly 7-8 percent over the same timeframe, driven by new enterprise users, said Joseph, and at the very high-end of the market exascale systems are driving growth.
Other major growth areas are the use of clouds for running HPC workloads (17.6 percent five-year CAGR), and storage (8.3 percent five-year CAGR).
End users spent $4.3 billion running HPC workloads in the public cloud in 2020, according to Hyperion, and that number is projected to more than double over the next four years, reaching $8.8 billion in 2024. That’s more than 2.5 times the growth rate of the on-prem HPC server market.
When public cloud is added to traditional on-premises purchases, the 2024 market forecast jumps to $47 billion. The public cloud spend makes up the second largest segment (with an 18.7 percent share) next to servers (40.5 percent share). Storage is a close third (with a 17.2 percent share).
HPC in the cloud hit a tipping point in 2019, Joseph said, where the numbers came in $1 billion higher than Hyperion had forecasted.
Now Hyperion is predicting a second tipping point for HPC cloud spending, sometime in the 2022-2023 period, Joseph said.
The analyst firm lists several potential big drivers and notes that if only two or three happen, there could be another billion dollar surge. Cited drivers include dramatic in-cloud performance improvements, increased ease of use, cost effectiveness/transparency, a significant uptick in AI workloads, and/or an emerging workload that is particularly well-matched for cloud.
Storage will likely also see major growth, Joseph said, driven by big data, AI and machine learning as well as larger traditional simulations.
“It’s possible that the storage growth numbers will actually be larger than what we’re showing,” said Joseph. “It all depends on how quickly the AI, machine learning and big data applications are successful and are applied in more areas. At some point in time, we expect storage to go through a tipping point, much like we saw the cloud, where there’s going to be a major increase in usage.”
The dawning era of exascale is another growth driver.
Exascale and near exascale systems are nearing readiness in China, EU, the UK, Japan, U.S. and other countries. Four-to-six systems per year, representing on the order of $2 billion a year, will be coming online starting within the next 12-24 months, according to Hyperion’s research. By 2026, Hyperion forecasts the total (cumulative) value will reach $10 billion to $15 billion.
“We are expecting 2022 to 2024 to be strong growth years, driven heavily by the exascale systems coming onboard, AI and HPDA, as well as HPC spending in the cloud,” said Joseph. “The exciting part to me is all the new technologies are showing up, whether it’s in the processor side, the hardware, the software, new storage approaches, memories, and everything. So it’s a very exciting time in the marketplace right now, providing many different choices for users.”
Hyperion has also updated its ROI report that examines investments in HPC projects and their financial and innovation returns. Based on 763 successful (revenue-generating) HPC projects, Hyperion’s model shows $507 in revenue was generated per dollar of HPC invested. On the profit and cost saving side, Hyperion saw $47 generated per dollar of HPC invested. For the next phase of the ROI study, Hyperion is working on including the projects that were not successful, so that it’s possible to determine the ROI generated by a given supercomputer or datacenter over time.
The ROI reports and the methodology (as well as the raw data) are available on Hyperion’s website. “The entire database is available to be downloaded by anyone in the community. So for example, if you want to use a different calculation, a different economic model than what we’ve used here, you’re able to do that,” said Joseph.