In March 2020, the near-term future of the high-performance computing market was suddenly thrust, like so many things, into turmoil and uncertainty. Intersect360 Research, which had previously forecast seven percent year-over-year growth for the market, reversed course, projecting anywhere between zero growth and negative 12 percent. Nearly 20 months later, a lot has changed: after avoiding the worst fears of the pandemic’s impacts on the market, HPC is rebounding – but, as Intersect360 shared in its November market update, the picture is far less rosy than it was in June.
Despite Intersect360’s best-guess prediction that the pandemic would bring the HPC market down by six percent in 2020, it actually ended up much closer to the flat-growth end of the prediction range, with just a 0.2 decrease for 2020 relative to 2019 ($39 billion to $38.9 billion). That led Intersect360 to project, in June of this year, that the HPC market would rebound to $60.2 billion by 2025, with 2021 showing a post-pandemic rally of HPC sales as purchases delayed by the pandemic poured in amid vaccinations and reopenings. All this, Intersect360 thought in June, might add up to something like a 16.4 percent year-over-year growth for 2021 relative to 2020.
“A lot can change in just a couple of months,” said Addison Snell, CEO of Intersect360, in this update.
A long road to recovery
To be fair to Intersect360, that June prediction had arrived with significant caveats: first, the firm had noted, it assumed that “most economic activity returns to ‘normal’ by the end of 2021.” Of course, that imagined world didn’t exactly materialize.
“In June, I personally was feeling more optimistic about, hey, vaccinations are underway, we’re getting back toward more global ‘normal’ economic activity,” Snell said. “The reality here in November is that vaccinations and vaccination efforts have been uneven around the world, both due to global availability of vaccines and also due to people optionally saying ‘I don’t want a vaccine.’ The Delta variant has also become a different sort of reality in terms of ‘what’s the efficacy,’ and that has stalled a lot of rebound of certain types of economic activity.”
Entry- and mid-range systems, Snell explained, continue to experience more impacts from Covid delays or cancellations, with larger systems faring better. But in terms of Covid, regional variability matters, as well, with regional caseloads causing significant fluctuations.
The pandemic has also brought about – or at least, severely exacerbated – another market hindrance: the ongoing supply chain disruptions that have persisted since this summer. In June, Intersect360 had said that it was assuming that “microprocessor shortages will not be a significant damper on the 2021 HPC market from the supply side.”
“Today, it’s clear that they are having an effect,” Snell said. “What we’re working on now is ‘what’s the magnitude of that effect?’”
“Several users have reported to us longer lead times in ordering, longer lead times in delivering,” he continued. “This, again, is possibly concentrated more on entry-level and mid-range systems but not restricted to any particular vendor or region.”
All this adds up to bad news – at least, relative to what seemed like a light at the end of the tunnel in June.
The light in the middle of the tunnel
“We’re prepared to give a range,” Snell said. “That we’re not going to hit 16 percent; that we are still going to have year-over-year growth, but that’s coming from a very poor 2020. Our revised guidance is that our growth is going to be in the range of six percent to twelve percent year-over-year[.] No change in the five-year forecast at this time, but I’d say that if we’re on the low end of this range – more on the six percent side – that probably indicates more sales that are lost than deferred and we will start seeing deleterious effects even in the out years in this forecast, we’re just not prepared to release a model on that at this time.”
Where the growth lands in that range is, this year, particularly important, Snell explained: six percent would indicate effectively no bounceback from the pandemic; anything above seven or eight percent would reflect some bounceback, or at least a normal growth rate; “but even at 12 [percent], we’re not recapturing as much as we essentially lost in 2020.”
Still, even within this gloomy outlook, there are winners and there are losers. For instance, Snell said, not all vertical markets that intersect with HPC are rebounding at the same rate: while transport-oriented sectors like automotive and aerospace, for instance, were suffering due to lower travel rates and chip shortages, other sectors – most notably, biomedical – experienced strong growth.
And, of course, another winner soared above the competition: the cloud sector, which Intersect360 forecast in June to continue to grow at over 20 percent compound annual growth rate. Snell attributed this “very heavy growth in cloud computing” simply: “cloud,” he said, “does well with uncertainty.” And as the uncertainty of the pandemic has persisted, cloud HPC’s growth has shown no signs of slowing down. Still, even with that strong growth, 84 percent of Intersect360’s survey respondents said that their operations were still all on-prem, with only one percent reporting all-cloud computing.
One other big factor looms over the HPC market in 2021: whether or not HPE recognizes the revenue from the Frontier exascale system – the first of its kind in the U.S. – that is currently being installed at Oak Ridge National Laboratory. The revenue from that one system will move mountains for the HPC market this year, one way or another.
For the rest of the market, we can only wait and see. Intersect360 is particularly waiting for the first earnings reports from major firms that reflect periods after August, when supply chain issues began accelerating and vaccinations slowed.
Python slithers to the top. “Python is really now well-established as a dominant HPC programming language,” Snell said. “My old standard that I’ve been saying for decades is that … HPC basically is 25 percent C, 25 percent C++, 25 percent Fortran and 25 percent everything else, and the only thing that changes is the mix of everything else. That’s really no longer true. … Python is almost equal to the combination of C and C++.” Further, Snell said that there was more active development in Python than there was in Fortran.
Massive scalability isn’t as pervasive as one might think. “Big computing isn’t as big as we think,” said Dan Olds, chief research officer for Intersect360. “Half of typical jobs are on one to four nodes. Half!” Further, around three quarters of typical jobs scale to 16 nodes or fewer. Maximum scalability shows a similar picture, with 55.8 percent reporting maximum job scalability of one to 16 nodes.
AI and ML link to more and more HPC processes. More than a third of survey respondents said that they had integrated AI or ML into their HPC applications for steering or optimization, while more than half use ML for pre- or post-processing. Respondents reported that the most common frameworks were Theano, Scikit-Learn and Spark-Mlib, which Snell and Olds both commented went against their expectations (which included Pytorch, higher-ranked Spark and so forth).
To read more of Intersect360 Research’s November update, click here.