This year’s fantastic Supercomputing 2023 was back in full form. Attendees seemed to be glad that the show was back in Denver, which was a preferred destination over last year’s show in Dallas.
Horsepower remained a top topic, with GPUs, AI, and other hardware central to discussions. Software optimization was also interesting, with hardware manufacturing reaching its physical limits.
Record audiences came to the show, topping 14,000, beating the previous high in 2019.
Here were some observations from the show floor:
Diversity: Among the sparkling tech, inclusivity is a top theme at SC23. The keynote speaker, celebrated physicist Hakeem Oluseyi, challenged the savvy audience to treat people equally and fairly and to mentor and help others. He masterfully mixed Einstein’s Theory of Relativity with how mentors helped him through a troubled life to become a prominent scientist eventually.
On the second day, kids, too were roaming the floor. This policy was a stroke of genius to introduce a much younger audience to concepts of supercomputing.
“Do you know NASA?” a mother asked her daughter.
“Yeeeeahhhh, Mom,” the daughter replied.
And off the mother and happy daughter walked into the booth. It was heartwarming to see such creative engagement of future minds.
Freebies: Attendees were walking around with fabric bags, filling them up with freebies offered by companies. There were the usual paraphernalia — t-shirts, pens, and the like. I picked up one freebie, a stress ball, from one booth (too much to write!).
There was plenty of coffee flowing, and many had baristas making cappuccinos and lattes. I stood in line for my caffeine fix at the HPE stall but abandoned it after I was called for an interview. Apparently, the best coffee was being served by King Abdullah University of Science and Technology, which made a big splash on the floor.
Program managers: If you are looking for a career in supercomputing, being a program manager could be an option. I spoke to program managers at the Purdue and TACC booths, and both universities were looking to fill empty positions.
Program managers fill an essential role of connecting scientific researchers to personnel handling computing resources. The role is especially becoming more relevant with a wide range of accelerators available for AI and scientific computing.
For example, the Purdue program managers take in computing requests for scientific research and connect the researchers to hardware experts. The hardware experts determine the type of compute needed and assign the supercomputing resources.
A program manager’s ability to communicate is paramount in this role, and a skilled program manager who can understand science and computing could excel in this role.
Cooling solutions: What’s a GPU without liquid cooling? Cooling companies staked their claim on the SC 2023 show floor with GPUs and AI chips expanding. Expect more at next year’s show as AI, accelerated computing, and the number of GPUs multiply.
Bill Gropp, director of the National Center for Supercomputing Application (NCSA), identified the number of cooling companies on the floor as a trend.
Gropp also said that cooling solutions mitigate but do not solve an emerging problem of AI systems consuming much power to achieve results.
Alternative ways to cut energy from supercomputing are on their way. A switch to mixed-precision computing could dramatically cut power consumption. The DoE is also looking to wrap AI in between scientific computing tasks for faster time to science.
Search for Nvidia alternatives: Nvidia was like the Republican debates: the leading candidate decided not to appear. Nvidia did not splash its cult GPUs and let its software and system-building partners do the talking on their behalf. They did have a small booth off to the side.
But Nvidia’s opposition had a rich presence on the show floor. Groq, Cerebras, SambaNova, and NextSilicon were hyping their accelerators as a get-out-of-Nvidia-jail-card. The companies argued that using Nvidia meant facing the guillotine of vendor lock-in.
The specialized AI chips are catching on as research chips with DoE labs, and universities are also open to trying out the chips. NCSA, for example, is working with SambaNova on chips. Cerebras chips are already being used for training.
Nvidia also got some bad news about Microsoft developing its own AI chip, which was announced in Microsoft’s clashing Ignite event. All in all, Nvidia had a very good trade show, and has a clear lead until Intel or AMD come up with serious AI chip competitors.
Moore’s Law, please die: Moore’s Law scaling is dead. Top500 organizer Erich Strohmaier said so without batting an eyelid.
For years, Intel has twisted the definition of Moore’s Law to suit its business plans. It’s better that Intel bury the 1965 paper instead of redefining it based on some obscure fact in the paper.
Zettaflop: Intel predicted a zettaflop supercomputer – or 1000x faster than exaflop – by 2027. But Top500 organizer Erich Strohmaier said supercomputing will barely hit ten exaflops by 2030. That is partly because Moore’s Law scaling has ended.
CPUs and RISC-V: The overwhelming interest in GPUs buried any interest in CPUs, but the upstart RISC-V instruction set architecture was popular with attendees.
To explain it simply, RISC-V is an open alternative to proprietary x86 and ARM, and companies can patch together their own accelerated chip.
But RISC-V is nowhere close to the latest x86 and ARM-based chips on speed or software. The idea of collaborating openly on open chip designs is driving some enthusiasm behind RISC-V.
Significantly, the U.S. government has given instructions to use RISC-V processors to simulate the next-generation sparsity supercomputer, in which processors are closer to localized data. The EU has built a homegrown accelerator based on RISC-V called EPAC.
The European Processor Initiative (EPI), which is funded by the EU to create domestic chips, wanted to develop a RISC-V server CPU. However, a representative said the project was on the backburner as the EPI abandons in 2025, unless they get more funding from the EU. Until then, most of the effort is focused on EPAC.
Chinese company Sophgo came to Denver to sell its 64-bit RISC-V server chip. At the RISC-V Summit earlier this month, Ventana Microsystems introduced a server chip that matches up with x86 and ARM on features including chiplets and the latest manufacturing.
Keep your eyes out on RISC-V.
Rich getting richer: The performance of the top 7 systems in Top500 equals the performance of the remaining 493. The supercomputing system Aurora could top an Exascale by the next Top500 listing, and soon after, the two-Exaflop El Capitan should take over the top spot. The total performance of Top500 could be concentrated in even fewer systems, especially given the number of systems submitted to Top500 has progressively declined since 2017.