China's Supercomputer Returns to World No. 1 Without NVIDIA and AMD

The Japanese-language publication PC Watch published an article on June 25 titled, “World’s Fastest Supercomputer Achieved with CPU Alone—No NVIDIA or AMD.”

On June 23 (German time), the TOP500 list of supercomputers was refreshed, with China’s supercomputer "LineShine" achieving a peak performance of 2.19 EFLOPS using only CPUs, securing the top spot globally. This marks the first time since 2017, when the "Sunway TaihuLight" claimed the title, that a Chinese supercomputer has reclaimed the world’s number one position after nearly nine years.

LineShine is installed at the Shenzhen Supercomputing Center (NSCS). Its standout feature is becoming the first TOP500 system to exceed 2.1 EFLOPS using only CPUs without any GPU assistance.

LineShine employs a proprietary 304-core LX2 processor running at 1.55 GHz, interconnected via a custom "Qilin" interconnect technology, totaling 13.79 million cores. The operating system is Qilin OS. With a power consumption of approximately 42.2 megawatts, it achieves an efficiency of 52.07 GFLOPS/W. It also ranks first in the HPCG (High Performance Conjugate Gradient) benchmark with a score of 22 HPCG-PFLOPS/s. In the HPL-MxP mixed-precision benchmark, it ranks fourth with 7.92 EFLOPS/s, demonstrating its design is optimized purely for CPUs without specialized low-precision accelerators.

The self-developed CPU "LX2" within the LineShine system integrates a matrix acceleration unit, enabling efficient coordination across multiple computing modes, combining high-speed computation with exceptional flexibility. It is also the first domestically manufactured CPU in China to adopt HBM, offering bandwidth ten times higher than traditional CPUs. The network further leverages proprietary "Qilin" technology, supporting large-scale connections with up to 2 million ports and 100,000 nodes. The storage layer adopts a hierarchical architecture, balancing high performance with massive capacity. Additionally, a full-stack software stack has been developed to make hardware easier to deploy directly from application stages. The system uses 100% liquid cooling, achieving an efficiency of 51 GFLOPS/W.

Since its launch, the system has supported applications including atmospheric and oceanic research, engineering simulations, materials science, drug discovery, neuroscience, scientific AI, and large-scale model inference. It has achieved an average scalability of 84.4% in mixed-precision computing, workload diversity, and large-scale parallel environments.

With LineShine claiming the top spot, the previously leading U.S. system El Capitan drops to second place. Meanwhile, Italy’s “HPC 7,” utilizing the same HPE Cray EX255a and AMD Instinct MI300A architecture as El Capitan, enters the rankings for the first time at sixth place. Japan’s “Fugaku” ranks ninth.

Original source: toutiao.com/article/1868943261414420/

Disclaimer: The views expressed in this article are those of the author(s).