Hyperion Research delivered its annual ISC 2021 HPC market update and this year’s version comprised a 100-plus slide survey of the landscape presented by six analysts. We won’t cover all 100-plus slides. Bottom line: Despite the pandemic, the broad 2020 HPC on-premise market (servers, storage, et al.) edged very slightly up (~1.1 percent) to $27.75 billion, while HPC in the cloud jumped (~10 percent) to $4.3 billion. Hyperion is calling for a slight dip next year before returning to growth.
“The year actually turned out positive and it was positive because the Fugaku machine, which was on the order of a billion dollars, actually changed the market from [being] down by 7-to-8 percent to be up just slightly,” said Earl Joseph, CEO, Hyperion. Besides Fugaku, which came online during the pandemic and has remained atop the Top500 list since, there was server strength at the high-end driven by exascale system sales and other large systems. The workgroup category declined the most, noted Joseph, hit hard by the migration of HPC to the cloud in a trend that is likely to persist.
Notably, the surge in sales of extremely large supercomputers, of which Fugaku is an example, has prompted Hyperion to reorganize its supercomputer segments. It will now track three sub segments going forward: Leadership/Exascale-class systems; Supercomputers-Large ($3M and up); Supercomputers-Entry Level ($500k to $3M). Note that the lower boundary for the overall supercomputing division ($500k) has not changed.
“There [are a couple] of reasons we’re doing this. The exascale/leadership class machines are on a very different trajectory than the rest of the market. First of all, just the sheer size of these systems with their price tags; we’re now talking over $300 million, many at $500 million, [some] that actually hit a billion dollars. You can also see over the next five years [that their] growth trajectory is very different at 23 percent. It’s a substantial change so look in the future for us splitting the market this way,” said Joseph.
In at least one sense the new labels make sense; they more accurately reflect the power of the systems in the subcategory, i.e.an entry-level supercomputer is just that, and may help clarify some of the previously opaque contents of Hyperion’s supercomputer category.
Capturing the full Hyperion presentation is beyond the scope of this article; moreover many of the numbers (understandably) haven’t changed since Hyperion’s last market update presented in May at the HPC User Forum, which HPCwire covered (see Hyperion: HPC Server Market Ekes 1 Percent Gain in 2020, Storage Poised for ‘Tipping Point and Hyperion Offers Snapshot of Quantum Computing Market).
To make finding what you are looking for inside the latest update easier, Hyperion has divided its lengthy ISC presentation into distinct, video segments by category and posted links to each segment. Those links are included at the end of the article.
Hyperion is currently forecasting on-premise servers of all types to grow 6.8 percent through 2024 after a slight dip in 2021. Interestingly, within the server category, Hyperion says dedicated HPDA/AI servers revenue are growing 4x faster than overall on-premise servers.
While server sales represent most of the HPC market (42 percent according to Hyperion), the storage category has long occupied the number two slot (17 percent in 2020). Hyperion senior analyst Mark Nossokoff made a case for storage becoming even more important. Driven by the familiar data deluge throughout computing, not just in HPC, also by the growth of data hungry AI technologies, and lastly by efforts to co-locate memory closer to compute, storage is poised to carve out an even bigger piece of the HPC compute pie he argued.
“Storage continues to be the largest growth statement of the various HPC broad market areas, with a forecast of 8.5 percent five-year growth rate, culminating in $8 billion of revenue in 2024,” said Nossokoff. “This represents roughly 20 percent of users’ overall spend on premise HPC infrastructure. For every dollar spent on on-premise servers, roughly 40 cents is spent on storage. HPC storage in the cloud is growing faster, more than twice the growth rate of on-prem storage spanned across the forecast period, representing roughly one-third of user spending on HPC resources in 2019. Cloud storage is forecast to be almost $3 billion in 2024.”
As evidence of storage’s growing importance in HPC, Nossokoff cited specifications for several forthcoming systems (see slides below).
“Several of the Department of Energy’s new and upcoming leadership machines have shared more detail relative to their respective storage architectural elements. NERSC invested quite heavily in Perlmutter to optimize Lustre for simultaneous heterogeneous, large block sequential and random small block access for what is anticipated to be one of the fastest AI-based HPC machines when it’s fully up and running,” said Nossokoff.
“Lawrence Livermore provided some insight into El Capitan storage subsystem. Dubbed Rabbit the subsystem will simultaneously support both shared file system access and a low latency direct access path from the compute nodes. And while not exclusively storage related, the flux resource manager is being designed to have locality awareness of resources, and optimally allocate them at a more granular level than traditional resource managers. Lastly, Oak Ridge (National Lab) recently shared the storage specs of Frontier anticipated to be the United States’ first exascale computer. The specs are impressive from all perspectives,” said Nossokoff.
Nossokoff noted Samsung’s decision to move support of persistent memory from NVMe to CXL and said, “This reflects a potential accelerated shift of CXL adoption within HPC for CPU, GPU, memory and storage hierarchy,” adding “Consumption and utility models are emerging, as evidenced by HPE’s support of HPC workloads on Green Lake and VAST’s shift from a hardware infrastructure vendor to a software company.”
File systems also continue to evolve. “In addition to the Lustre innovation already mentioned, Seagate announced the release of CORTX, an object archive file system, in conjunction with its live strategy of on-prem storage infrastructure and cloud storage utility. Another interesting file system development was HPE’s support of IBM Spectrum scale, reflecting the growing adoption of HPC-enabled AI infrastructures in the enterprise.”
Last but certainly not least in the HPC world, he said, were announcements from several large HPC sites sharing their strategies of increased leverage of the cloud for HPC storage. “Almost four exabytes of storage is expected to be deployed as part of the UK Met-Microsoft deal for addressing the former’s next-gen requirements for weather climate research and storage query and retrieval of data within its active Data Archive. Rounding out this non-exhaustive cloud storage activity is Riken’s announcement of partnering with Oracle for its cloud storage needs on Fugaku.”
Two Bonus Slides
Video Links to the Hyperion Research HPC Market Update During ISC21:
- Earl Joseph: ISC21 HPC Market Update: https://youtu.be/RHktfkY9JqU
- Alex Norton: New Trends in Using Clouds for HPC Workloads: https://youtu.be/Cgo02L8LGqc
- Steve Conway: HPC-Driven AI Market Trends: https://youtu.be/hYvo1U4CAhQ
- Alex Norton: Highlights of the Continued Growth for HPC-enabled AI: https://youtu.be/76vmknuXTOg
- Mark Nossokoff: Storage Market Update: https://youtu.be/cP1lvAkjE7g
- Bob Sorensen: Quantum Computing: Finding Its Place in the Advanced Computing Sector: https://youtu.be/EbzBXDMySeo
- Melissa Riddle: Application Software Study Results: https://youtu.be/Qzw0H21iyc4
- Bob Sorensen: A Quick Tour of Exascale Architectures: https://youtu.be/z-w_yFs2jmU
- Earl Joseph: Conclusions: https://youtu.be/Tm1-NKxy854