Every year, SC has a theme. For SC22 – held last week in Dallas – it was “HPC Accelerates”: a theme that conference chair Candace Culhane said reflected “how supercomputing is continuously changing the world by accelerating the process of reaching solutions.” At SC22, just before the show floor opened, the HPC Accelerates Plenary brought together leaders from a range of sectors to explore “The Many Dimensions of HPC Acceleration.” The discussion, moderated by Addison Snell, CEO of Intersect360 Research, sought to explore how supercomputing both literally and figuratively accelerates.
Participating in the discussion: Jim Cherry, associate director and chief of research technologies at the National Institute of Allergy and Infectious Diseases (NIAID); Thomas Schulthess, director of the Swiss National Supercomputing Centre (CSCS); Shubho Sengupta, an AI researcher with Fundamental AI Research (FAIR) at Meta; and Gina Tourassi, director of the National Center for Computational Sciences (NCCS) and the Oak Ridge Leadership Computing Facility (OLCF) at Oak Ridge National Laboratory (ORNL).
“There’s literal acceleration, I suppose – computation makes the application run faster,” Snell said in opening the discussion. “But beyond that, are we accelerating scientific discovery? Are we accelerating human endeavor? There are a lot of lofty things we can accelerate.”
“When you strive to be at the bleeding edge of high-performance computing, the ultimate goal is to accelerate computation for some of the most pressing and challenging problems,” Tourassi agreed. “And certainly over the past few years, we have all experienced all the existential threats that humanity faces – starting with climate change, resource scarcity and pandemics. And when you think about what we have faced as a society in the past three years, all of these threats acting together, you realize how we need to bring these resources and accelerate innovation even faster.”
Tackling these problems, Tourassi said, requires us to develop different understandings of complex systems – the physical, the biological, the ecological and the social. “And HPC always has been at the core of helping us model and understand these systems and these interactions,” she continued, adding that more acceleration was necessary: “Honestly, I feel that we are running out of time with many of these challenges.”
Schulthess stressed how the enormity of those challenges was transforming the necessary solutions. “[Because of] these problems that go beyond what you do in one country, we’ve been forced … not to think just about computing, but to think about the data sources, how to organize the data across different countries,” he said, saying that user communities were facing “data volumes that you cannot manage as individual users any more” along with more data variety, distribution and characteristics. “It is obvious that we have to combine the collection of data, the processing of data and the analysis and the scientists who take interest in what information is in that data,” he said.
To that end, Cherry emphasized the value of partnerships and collaborations for accelerating HPC and tackling challenges – value that he suggested was growing as data became increasingly overwhelming. “The mindset is changing,” he said. “20 years ago, it was ‘get as many postdocs in the laboratory as possible, and get as much as data out there now’ – these machines generate such a tsunami of data, it’s ‘what does this data tell us?’”
Highlighting a range of partnerships involving NIAID, Cherry said that staff-sharing had been essential to accelerating research. “At the end of the day it really allowed us to move science forward much more rapidly at rate,” he said, “because when you only have one informatician, or one person that does a lot of the infrastructure and HPC, you’re gonna slow down. It’s gonna come to a grinding halt.” Leveraging the economies of scale, he said, meant that “everyone is able to get their answers more rapidly, more efficiently.”
“For us, collaboration is a guiding principle,” Tourassi agreed. ORNL, she said, starts with vendor collaboration – necessary to jointly design and deploy bleeding-edge systems – but also shares practices among its fellow national labs, works with analogous international entities and engages with a broad range of user communities, each of which has different needs.
Sengupta, for his part, said that collaboration worked differently for a lab that existed inside of a corporate entity – even though FAIR, he said, liked to think of itself as an academic lab. Still, he continued, FAIR linked itself with the leading universities wherever it put down roots, leading to deep partnerships with the universities in Paris, UC Berkeley, University College London and more. Sengupta also said that FAIR works to open-source datasets and collaborate on code frameworks, such as PyTorch (which FAIR originated).
But Sengupta also discussed another barrier to acceleration. AI researchers, he said, were working to close the loop between building, training, testing, deploying and rebuilding models as swiftly as possible; but, he said, he was troubled by the competition to simply build the largest model when all the models – namely, large language models (LLMs) – “use the same-ish technique.”
“Nowadays,” he said, “everything is really one large transformer … Is this it?” Radical ideas, he added, came from the fringes – and he wanted to know what was beyond the transformer paradigm.
Tourassi also commented on what she called the “move, move, move” narrative surrounding HPC-driven acceleration, citing – like Sengupta – the use of LLMs as a “magic bullet” for AI deployment.
“It is difficult to quantify acceleration of science … but I believe that we all feel that everything moves a lot faster than ever before,” she said. “And if we move too fast, I believe that we tend to harden our position with certain ideas, and we forget to apply the same level of scientific rigor that we would have applied if we were not under that continuous pressure of competition, innovation – that competitive advantage.”
She added, though, that competition was not always a bad thing: “Historically, competition and collaboration both fueled scientific research and productivity. You need competition to drive innovation. You need collaboration to accelerate it.”
Allusions to the negative impacts of unscientific policies and cumbersome bureaucracies pervaded the plenary, coming to a head when Snell asked the panel about the role HPC had in accelerating social thought in the face of disinformation and the rejection of science.
“In the end you need a system that respects the individual,” Schulthess answered. It was essential, he said, to make sure that acceleration reflected social interest. “If you accelerate too much you end up losing the people,” he said, saying the community needed to “bring society along” in its accelerated endeavors. “We do this by accelerating things they care about.”
Cherry agreed, adding that it was important in his field to bring clinicians and basic researchers along with any AI and HPC researchers. “They’re gonna spend time with the clinicians, understand how we can use this tool to bridge your scientific questions. So then you are bringing that society … along with us.”
The panelists also identified common ground in the area of isolated software and data. Snell said that, as an analyst, he was tracking an increase in nationalism and the siloing of HPC and research efforts. The other panelists sounded off on the detriments of this siloing (“Proprietary software doesn’t work,” Schulthess proclaimed at one point, earning a raucous audience reaction).
Tourassi, meanwhile, highlighted areas for improvement: “I would love to see our HPC community work more methodically on finding solutions in the space of privacy-preserving computing and federated learning,” she said. “Every time you have policy-related issues, they move incredibly slowly, and technical solutions are easier to get than policy solutions.”