ASC23: Tenth Anniversary Edition – Still the Toughest

By Dan Olds

The tenth edition of the Asian Supercomputing Challenge took place last week in Hefei, China, and it was tougher than ever. 24 university teams (20 on site, 4 remote) competed for five days to see which team could optimize real-world HPC/AI applications and benchmarks on systems they configured themselves. The only limit on the student teams was a 3,000-watt power cap that couldn’t be exceeded during their application runs.

The ASC competitions are the most demanding student cluster competitions in the world today. Why? It’s the applications. The ASC slate of apps, plus the pressure of the judging interview, is always a mind-bending adventure for the students and this year is no different. Here’s what the students had to contend with at ASC23:

Benchmarks: HPL and HPCG, which are staples in every student cluster competition. These are the first applications in the formal competition and serve as a relatively easy first step towards the much more intense upcoming tasks. Every team in ASC23 turned in both an HPL and HPCG score, which is admirable.

YLLM: this is a large language model which several of the teams thought might be the most challenging task in the competition. The teams are being asked to build a language model from a 100GB dataset provided by the organizers. But there’s a catch: they have to build two models, one generating 1.157 billion tokens and one generating 17.888 billion tokens.

DeepMD: adds AI analysis to molecular dynamics models such as those generated by LAAMPS or GROMACS (plus others). This radically increases the efficiency of molecular modeling. Student were tasked with modeling water and copper, along with other cases. DeepMD won the 2020 Gordon Bell prize at SC20 and is in wide use through the HPC world. We took a deep dive into DeepMD in our interview below with Linfeng Zhang, Chief Scientist and Chairman of DP Technology.   

WRF-Hydro: if you want to predict flooding, snowpack, rain, or anything else having to do with water and the earth, WRF-Hydro is what you need. It’s open source, it scales, and it can model just about anything water/weather related, what more could you want? It’s also a difficult application to run and optimize, judging by the results from ASC23. While most of the teams managed to turn in a result, the top scores were around seven points (out of 18 max), but the median score was only 2.79. A painful app for the student teams.

Mystery Application – FVCOM: is designed to analyze and model coastal ocean water circulation plus other things such as salinity. The mystery application in a cluster competition is always a wild card. They can be well known applications with lots of documentation or they can be obscure codes that students will struggle to compile and optimize. FVCOM is falls somewhere in the middle, but still turned out to be a difficult challenge, much to the student’s dismay. Average score on this one was only 2.14 points out of a maximum of 18, making it the lowest scoring application in the entire competition.

Group Challenge: to mix things up and get the students working together, the organizers designed a group challenge that randomly matched up members from each university into teams who were then tasked with tackling a quantum simulation challenge. While the results from the group challenge weren’t factored into the overall results, there was a 20,000 CNY (about $2,800 US) prize pool for the winning set of students. This was quite a fiendish challenge, as it turned out. To learn more about it, we interviewed Pan Feng from the Shanghai Research Institute, a branch of the University of Science & Technology of China, and the designer of the task.

Judging Interview: in this final challenge, the students face a room full of HPC experts and have to give a presentation covering their competition experience. This includes their efforts and results on every competition application. It puts quite a bit of pressure on the students but plays a vital role in ascertaining just how much the students learned during the event. To give you a better understanding of the judging process, I interviewed long-time ASC judge Guan Yongqiang, Honorary Senior Associate Researcher of the University of Hong Kong. What followed was a delightful conversation about student cluster competitions and his 40+ year long experiences in HPC. It’s a lively and wide-ranging chat that I truly enjoyed.

Next up in our on-the-spot coverage we’ll give you a chance to meet the teams via the miracle of video. Stay tuned….