中国学生梦寐以求的设备在世界大赛中击败日本最强大的超级计算机  (ZT, Google 软件翻译）

2022 年 6 月 15 日，星期三，上午 5:30 根据最新的全球排名，由中国学生开发的小型计算机在解决与人工智能相关的重大复杂数据问题方面的表现优于日本最强大的机器。 日本的超级计算机 Fugaku 拥有近 400 万个 CPU 内核，使其成为有史以来第二大计算机。 DepGraph超级节点是武汉华中科技大学研究生作为“培训项目”启动的，拥有128个核心。 您对来自世界各地的最大话题和趋势有疑问吗？通过 SCMP Knowledge 获得答案，这是我们的新平台，由我们屡获殊荣的团队为您带来的解释器、常见问题解答、分析和信息图表。 但根据国际超级计算发布的年度 Graph500 排名，DepGraph 在解决单源最短路径（SSSP）问题方面的速度几乎是 Fugaku 的两倍，这是一个影响人工智能在广泛领域的性能的难题。本月初召开会议。 “感觉就像一场梦，”《中国科学报》周一在一篇报道中引述一年级研究生沉千歌的话说。 沉和她的队友的平均年龄是24岁。 数学家经常使用图表来描述关系。最简单的图形可以只包含两个点，它们之间有一条线。 更复杂的图表已经应用在很多领域，比如拥有大量上市公司的金融市场、全球社交媒体平台或战争游戏。 人工智能可以检测隐藏的关系或在复杂的、不断变化的图表中发现进化模式。 但训练和学习过程通常涉及大量计算。即使是一些看似简单的任务，例如寻找两个点之间的最短路径，也对计算机提出了巨大的挑战。 武汉团队表示，他们发现了一个瓶颈问题，可能会严重影响这类工作的计算机。 根据该团队去年提交给 IEEE 高性能计算机体系结构研讨会的一篇论文，超级计算机使用许多 CPU（中央处理单元）内核来同时执行许多计算任务。但是在处理与图相关的问题时，一个核心的计算过程通常取决于另一个核心的结果。 该论文称，这种依赖性扰乱了计算过程。大多数时候，核心要么等待，要么用另一个核心的新结果重新计算。 该团队表示，DepGraph 机器通过一种新的结构和软件解决了这个问题，可以通过减少依赖引起的混乱来将每个核心的性能发挥到极致。 “不要低估学生的能力，”计算机科学副教授、项目组顾问张宇说。 “这是第一次单台计算机在图形计算中击败一组计算机，”他补充说。 学生的导师金海教授说，这些年轻的研究人员与中国的高科技公司密切合作。他们最大的灵感来自于这个行业。 《中国科学报》援引他的话说：“问题肯定出在行业。这项研究满足了我国最迫切的需求。” 博士生兼团队负责人赵金表示，他们必须每两周向中国科技巨头华为技术公司报告一次进展情况。 “硕士和博士生可以接触到国家重点研发计划的课题，这让我们瞬间大开眼界，让我们与行业更紧密地融合在一起。”他说。 但赵说，他们也享有很高的自由度。 “我们自己找到我们的研究课题。导师不会告诉学生该做什么，而是在我们感兴趣的方向上支持我们。这非常令人兴奋，”他补充道。 根据 Graph500，Fugaku 在广度优先搜索 (BFS) 中记录了最高性能，这是另一个主要的图问题，其算法更简单但需要更多的原始计算资源来解决。 2015年建成的中国超级计算机神威太湖之光在该类别中排名第二，DepGraph在巨机中排名第九。 中国已经建造了一些世界上最大的计算机，其中包括一些可以与美国能源部 Frontier 相媲美的百亿亿次计算机，这是目前公开记录中最快的计算机。 根据公开信息，中国研究人员进行了前所未有的实验，例如在这些新机器上进行世界上最大的人工智能培训。 但中国当局对他们的表现保密。 据行业专家称，超级计算机国际排名的可靠性受到中国缺席的影响。 本文首发于最权威的语音代表南华早报（SCMP）

Chinese students' dream device defeats Japan's most powerful supercomputer in world contest
Wed, June 15, 2022, 5:30 AM
A small computer developed by Chinese students outperformed Japan's most powerful machine in solving a major complex data problem related to artificial intelligence, according to the latest global ranking.

Supercomputer Fugaku in Japan has nearly 4 million CPU cores, making it the second-largest computer ever built.

DepGraph Supernode, which was started as a "training project" by graduate students at Huazhong University of Science and Technology in Wuhan, has 128 cores.

Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.

But the DepGraph was nearly twice as fast as Fugaku in solving a single source shortest path (SSSP) problem, a difficult graph problem affecting the performance of artificial intelligence in a wide range of sectors, according to the annual Graph500 ranking released by the International Supercomputing Conference early this month.

"It felt like a dream," first-year graduate student Shen Qiange was quoted by China Science Daily as saying in a report on Monday.

The average age of Shen and her teammates was 24.

Mathematicians often used graphs to describe relations. The simplest graph can contain just two dots with a line between them.

More complex graphs have been applied in many areas, such as financial markets with a large number of listed companies, global social media platforms or war games.

AI can detect hidden relations or discover a pattern of evolution in a sophisticated, constantly changing graph.

But the training and learning process usually involved a large number of calculations. Even some seemingly easy tasks, such as finding the shortest path between two dots, imposed a huge challenge to computers.

The Wuhan team said they discovered a bottleneck issue that could severely affect computers on this type of job.

A supercomputer uses many CPU - central processing unit - cores to execute many calculation tasks simultaneously, according to a paper the team submitted to the IEEE Symposium on High-Performance Computer Architecture last year. But when handling a graph-related problem, the calculation process in a core often depends on the outcome of another.

This dependency disrupted the calculation process, the paper said. Most of the time the cores had to either wait or re-do calculations with new results from another core.

The DepGraph machine solved the problem with a new structure and software that could take the performance of each core to the limit by reducing the chaos caused by dependency, the team said.

"Don't underestimate the ability of students," said Zhang Yu, associate professor of computer science and adviser to the project team.

"This is the first time a single computer defeated a cluster of computers in graph calculation," he added.

Professor Jin Hai, a mentor to the students, said the young researchers had worked closely with China's hi-tech companies. Their biggest inspirations came from the industry.

"The problem definitely originated from the industry. The research meets the most urgent needs of our country," he was quoted by the China Science Daily as saying.

Zhao Jin, a PhD candidate and team leader, said they had to report their progress to Chinese tech giant Huawei Technologies every two weeks.

"Master's and doctoral students can get access to the topics of the national key research and development plan. This opens our eyes instantly and integrates us more closely with the industry," he said.

But Zhao said they also enjoyed a high degree of freedom.

"We find our research topic by ourselves. The supervisors do not tell the students what to do, but support us in the direction we are interested in. This is very stimulating," he added.

Fugaku recorded top performance in breadth-first search (BFS), another major graph problem with an algorithm that was simpler but required more raw calculation resources to solve, according to the Graph500.

Sunway TaihuLight, a Chinese supercomputer built in 2015, ranked second in that category, with the DepGraph ranking ninth among the giant machines.

China has built some of the world's largest computers, including some exascale machines that could rival the US Department of Energy's Frontier, currently the fastest machine in open record.

Chinese researchers have conducted unprecedented experiments such as the world's largest training for AI on these new machines, according to openly available information.

But the Chinese authorities have kept their performance secret.

The reliability of international ranking on supercomputers has been affected by China's absence, according to industrial experts.

This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice rep