The semiannual taking of HPC’s pulse by Hyperion Research – late fall at SC and early summer at ISC – is a much-watched indicator of things come. This year is no different though the conversion of ISC to a digital event forced Hyperion to simply issue the report rather than present the results at ISC. Top line: 2019 was a flat year (revenue) compared to a booming 2018, reported Hyperion, and 2020 is tumbling thanks to the COVID-19 pandemic, though how low it will go remains unclear.
The pandemic has made many things murky.
Steve Conway, Hyperion senior advisor, told HPCwire, “The Q1 (2020) numbers are almost locked, and it looks like Q1 is going to come in 15-to-18% lower than last year’s first quarter. It’s tending toward that 16% number now – not a disastrous departure from the prior year, but definitely down. Q2 is the one that we’re most curious about.
“What people have told us a couple of months ago was that orders weren’t being cancelled. Things are mainly getting pushed out and sort of the higher up you go in the price range, when you get to exascale-class systems, it doesn’t look like that development has taken much of a hit at all. They seem to have redoubled their efforts to, to make sure that they have the parts that they need, and so forth. The biggest hit is in the work group segment.”
The implication is strength at the high-end will offset some of the slippage below. Hyperion’s most recent pre-COVID 2020 forecast was for 8.7 percent growth (CAGR) in HPC to $14.4B. That won’t happen but how much of a decline won’t become clearer until Q2 numbers are final, said Conway. Hyperion reports supercomputing revenue will still grow. Below are Hyperion’s HPC market numbers issued in June and not yet revised down to accommodate pandemic effects.
One of the more interesting developments reported by Hyperion is Arm’s sudden rise in HPC – much of that fueled by big systems such as Japan’s Fugaku supercomputer which topped the recent Top500 List and uses a Fujitsu’s 48-core Arm A64FX SoC. There are three other Arm systems on the Top500: the A64FX Fugaku prototype at Fujitsu (#205); the new Fujitsu PRIMEHPC FX1000 A64FX system, Flow, at Japan’s Nagoya University (#37); and Astra, the Marvell/Cavium ThunderX2 installation at Sandia (#245), recognized as the world’s first petascale Arm system in November 2018.
Hyperion made its first “full forecast for Arm processors in its latest update.
As you can see in the slide above, Arm’s forecast jump in HPC is startling before flattening somewhat in 2023, presumably because of completion of a number of large system installations. With two major chip-makers now providing high-end Arm processors (Fujitsu and Marvell) and adoption by Cray of a Fujitsu Arm chip for its CS500 line (cluster supercomputer), Conway expects more big systems builders to embrace Arm. Many systems vendors have already flirted with Arm but now the stage seems set for wider adoption.
Geopolitical forces are also at work here, noted Conway. Countries increasingly see HPC as a necessary competitive tool and want more control over their computer technology. This includes owning their own big systems and greater control over computer component (e.g. microprocessor) supply lines. Current global trade tensions and pandemic-caused supply-chain interruptions have exacerbated the worries, said Conway.
Case in point: “There is an effort underway in the UK, a pretty solid effort now that they are no longer part of the European Union, to have an exascale system. That’s being led by three universities, Edinburgh and Leicester and Bristol. Bristol is where all the Arm development is taking place. They are already on second generation Cray system with Fujitsu arm 64 FX processors,” noted Conway.
HPC (FINALLY) TAKES TO THE CLOUDS
After years of steady but perhaps unspectacular growth, Hyperion reported 2019 is a tipping point for HPC-related cloud spending by users. For several years, cloud providers have been beefing up their HPC resources – GPUs, fast CPUS, HPC orchestration services, applications, AI tools, parallel file system support, etc. The investment is now paying off with wider adoption by HPC users.
“It’s mostly burst capability at this point. Still, there’s not a lot of cannibalization by the cloud. Mostly it’s things that the on-premises folks wouldn’t have been able to run before. But what’s really driven this most is not the workloads. It’s been the cloud services providers. You’ll remember in 2011 when NASA took its famous Magellan study, and basically concluded that clouds [then] were only good for embarrassingly parallel workloads and nothing else,” said Conway. That’s changed.
There are two big drivers said Conway:
- Market Size. “In the supposed golden age of HPC, in the early 1990s, the whole HPC market with everything thrown in – storage, servers, software, etc. – was worth about $2 billion worldwide. In 2019, [it’s] $28 billion, and we’re saying $43 billion in 2024, not counting HPC cloud usage, which brings it up to about $51 billion. That’s a decent sized market to go after. It made it more attractive for CSPS (cloud services providers) to pay attention and change their platforms.
- AI’s Ascent. “The other part is AI. It turns out, I guess not surprisingly, that at the forefront of AI, HPC is indispensable, whether you’re talking about automated driving or precision medicine or whatever, not the mainstream but at the forefront. And the forefront is where the mainstream is going to be,” said Conway.
STORAGE DEMAND REMAINS ROBUST AND AI ASCENT CONTINUES
Given the growth of data-intensive computing, especially AI’s proliferation, it is not surprising that storage requirements have also soared. Conway says storage has steadily become more complex and that HPC storage buyers have necessarily grown more sophisticated.
“It’s gotten very complicated because we’re sort of moving toward a very new model for HPC that affects every part of the architecture, but storage in particular. That model is the containerization of on premises HPC. We’re now looking at workflows that might have to go through 20 lightweight containers, each one of which has to quickly assemble the right compute storage software so the workflow can complete. That’s not the way things used to be it used to be, you know, this is your hardware, your software, and so forth, and everything was the same all the way.”
Hyperion also singled out accelerating AI development and adoption, no surprise here, and said AI will grow faster than the market as a whole.