ASC24

The preliminary results for the 2024 ASC Student Supercomputer Challenge (ASC24) have been announced. Over 300 universities worldwide participated, with twenty-five standout teams advancing to the finals. Among the finalists are renowned institutions such as Huazhong University of Science and Technology, Peking University, the Chinese University of Hong Kong, National Tsing Hua University, Friedrich-Alexander-University Erlangen-Nuremberg, and the National University of Cordoba. They will compete for prestigious awards including the Champion, Silver Prize, Group Competition Award, e Prize, and Highest LINPACK at the finals hosted by Shanghai University from April 9 to 13, 2024.

In response to the challenge posed by a large number of excellent teams, the ASC Organizing Committee, and Shanghai University have decided to accommodate 25 teams for the ASC24 Finals, setting a new record for the highest number of teams qualified for onsite finals since the inception of the ASC competition in 2012.

Among the finalists, Huazhong University of Science and Technology's team demonstrated an in-depth grasp of supercomputing knowledge and technologies. Through precise performance bottleneck analysis and innovative optimization methods, they achieved outstanding results. Their exceptional performance across all tasks secured them the top rank in the preliminary round.

The ASC24 finals feature a lineup of traditional powerhouses, including Peking University - the champion of the 10th ASC onsite finals; the Chinese University of Hong Kong - the champion of the 10th ASC virtual finals; University of Science and Technology of China; Shanghai Jiao Tong University; and Friedrich-Alexander-University Erlangen-Nuremberg. Additionally, there is great excitement surrounding the debut of teams from National University of Cordoba, Macau University of Science and Technology, and Southwest Petroleum University, marking their first appearance in the ASC finals this year.

The ASC24 preliminaries included two computational tasks: Large Language Model (LLM) Inference Optimization and Seepage Numerical Simulation. Numerous participating teams excelled, showcasing remarkable capabilities in exploring and innovating in supercomputing application analysis, performance optimization, and parallel strategy design.

In the Large Language Model (LLM) Inference Optimization Challenge, participating teams were assigned the task of building and refining inference engines using the widely used open-source LLM, LLaMA2. The main goal was to achieve high-throughput inference on sample datasets provided by the organizing committee. Teams contending for success were expected to showcase a thorough understanding and proficiency in common parallel methods applicable to LLMs. Furthermore, they were tasked with implementing various techniques to enhance the inference process.

The team from Huazhong University of Science and Technology developed an LLM inference engine that incorporated tensor parallelism, KVCache, and PagedAttention. Additionally, they implemented asynchronous pipelining in parallel to address the issue of incomplete loading of model parameters on a single server. Moreover, they deployed a distributed memory manager on each node to ensure load balancing and conducted numerous comparative experiments to identify the optimal optimization method. These efforts propelled them to achieve the highest score. On the other hand, the team from Friedrich-Alexander-University Erlangen-Nuremberg utilized a customized TensorRT-LLM engine for their cluster. Through various experiments, they determined that a batch size of 8 resulted in twice the throughput compared to a batch size of 1. By leveraging tensor parallelism and batched generation, they achieved good performance.

The Seepage Simulation Task in ASC24 was designed to delve into the intricate flow patterns and characteristics of multiphase fluids within porous media using the open-source software OpenCAEPoro. Participating teams were tasked with simulating the seepage of multiphase fluids, including oil, gas, and water, specifically in scenarios related to petroleum extraction, based on the dataset provided by the organizing committee. The challenge required teams to skillfully optimize large-scale parallel computing processes, with a focus on improving the computing performance and parallel efficiency of discrete algorithms.

The team from Qinghai University utilized common optimization methods, including function inlining, loop expansion, and operator fusion, after conducting a detailed analysis of the seepage simulation software OpenCAEPoro. They also implemented vectorization and memory access optimization to address computational bottleneck matrix assembly, resulting in highly effective optimization. Meanwhile, the National Tsing Hua University team conducted a comprehensive analysis of OpenCAEPoro, pinpointing the primary bottleneck for CPU parallelization. They leveraged optimized compiler flags and computing libraries to achieve the best performance.

About ASC

The ASC Student Supercomputer Challenge is the world's largest student supercomputer competition, sponsored and organized by the Asia Supercomputer Community with support from experts and institutions across Asia, Europe, and America. The main objectives of ASC are to encourage the exchange and training of young supercomputing talent from different countries, improve supercomputing applications and R&D capacity, boost the development of supercomputing, and promote technical and industrial innovation. The first ASC Student Supercomputer Challenge was held in 2012 and has since attracted over 10,000 undergraduates from all over the world. Learn more for ASC at http://www.asc-events.net/StudentChallenge/index.html.

To view this piece of content from cts.businesswire.com, please give your consent at the top of this page.

View source version on businesswire.com: https://www.businesswire.com/news/home/20240207108190/en/