China's LineShine tops supercomputer ranking with all-CPU architecture

SHENZHEN -- China's domestically-developed LineShine supercomputer has topped the latest TOP500 list with 2.198 EFLOPS of sustained double-precision performance, becoming the world's first supercomputer to sustain more than 2 EFLOPS on the High Performance Linpack benchmark, according to the National Supercomputing Center in Shenzhen, southern China.

The achievement marks a significant breakthrough in the convergence of supercomputing and intelligent computing, two fields that have traditionally used different hardware and precision requirements.

Traditionally, supercomputing has relied mainly on double-precision floating-point operations for scientific applications such as physical modeling and engineering simulation, while intelligent computing has used lower-precision or integer operations for model training and inference. In recent years, the two disciplines have been moving toward integration.

The conventional approach to hardware integration has been a heterogeneous CPU-GPU architecture, where the CPU handles scheduling and control while the GPU accelerates computation. However, this model suffers from high data-transfer costs, complex programming requirements and underutilized hardware resources.

LineShine takes a different approach, with its pioneering "Online Acceleration" all-CPU architecture. By embedding AI matrix acceleration units directly into its domestically designed processors, the system enables CPUs to efficiently execute AI tasks without relying on GPUs, reducing the CPU-GPU data-transfer bottleneck common in heterogeneous architectures.

Beyond the processor, LineShine incorporates large-scale innovations across networking, storage, system and energy efficiency. The result is a dual breakthrough in top-tier computing performance and broad real-world application deployment, providing a truly viable solution for the convergence of supercomputing and intelligent computing, said Lu Yutong, chief designer of LineShine and director of National Supercomputing Center in Shenzhen.

In terms of infrastructure for scientific and engineering intelligent computing, LineShine has already supported applications in fields including atmospheric and ocean science, engineering simulation, materials science, drug discovery, brain science, scientific AI and large-model inference, Lei Kai, deputy director of the Shenzhen Computer Federation, told Xinhua.

The TOP500 list is released twice a year. Chinese supercomputers have topped the list on multiple occasions. Tianhe-1 first claimed the No 1 spot in 2010; Tianhe-2 held the top ranking for six consecutive editions from 2013 to 2015; and Sunway TaihuLight topped the list four times from 2016 to 2017.