China’s LineShine Supercomputer Achieves 2 ExaFlops with 47,000 CPUs and No Dependence on Foreign Chips

China has recently introduced its cutting-edge supercomputer, LineShine, located in Shenzhen, promising an impressive computing power of over 2 ExaFlops.

LineShine: Poised to be the World’s Fastest Supercomputer, Exceeding 2 ExaFlops

Announced during a conference at the National Supercomputing Center in Shenzhen, the LineShine project aims to establish the fastest domestic supercomputer. This ambitious initiative will unfold in two phases and is expected to surpass 2 ExaFlops in computational performance. Currently, the title of the world’s fastest supercomputer belongs to the AMD-powered El Capitan, which achieves a remarkable peak speed of 2.8 ExaFlops.

Specifications of LineShine

Distinctively, the LineShine Supercomputer is a CPU-centric system designed to offer sustainable performance exceeding 2 ExaFlops. Its architecture utilizes high-efficiency CPUs, wide bandwidth memory, and rapid interconnects, all crafted to optimize processing capabilities.

Significantly, all components within the supercomputer are sourced from domestic manufacturers, ensuring complete reliance on local technology suppliers. Furthermore, LineShine is designed to be user-friendly, equipped with a comprehensive toolchain that includes compilers, debuggers, and tuning utilities, facilitating a practical ecosystem for developers.

A person in a suit is standing next to a display featuring a LineShine server rack with illuminated hardware components.

Phased Development

The construction of LineShine will occur in two stages. The initial phase will consist of a pilot verification system using 100 Huawei Kunpeng servers, amounting to a total of 12, 800 cores. In the subsequent phase, an industrial complex will host 1, 580 blade servers powered by x86 CPUs, culminating in a combined total of 101, 120 cores with a theoretical peak exceeding 10 Petaflops. The architecture includes a configuration of 16 servers arranged in a 4-way design and four 8-way servers.

This expansive system will incorporate 92 computing cabinets equipped with 47, 000 CPUs, alongside 36 network cabinets designed for massive scalability, accommodating hundreds of thousands of nodes. It features a million-port interconnect, solidifying its status as China’s largest supercomputing storage base. Additionally, it employs the world’s most extensive liquid cooling system, with 67 liquid-cooled storage cabinets, 428 storage nodes, and an impressive 10 TB/s of bandwidth.

Innovative Cooling Technology

The liquid cooling mechanism of LineShine is noteworthy, comprising secondary pipes that extend over 3, 214.7 meters and weighing in at 243.9 tons. This advanced cooling solution is vital for maintaining optimal operating conditions for the vast array of CPUs housed within the supercomputer.

A presentation featuring 'Supercomputing Phase II: A domestically produced exascale supercomputer system' with the 'LineShine Supercomputer Phase II in Shenzhen' is visible on a large screen, alongside technical specifications and a picture of the supercomputer.

Storage and Performance Enhancements

Emphasizing its outstanding storage capabilities, LineShine is engineered to provide a staggering 650 Petabytes of total storage capacity. Its innovative Fusion Architecture integrates simultaneous multithreading accelerators that significantly enhance matrix and mathematical processing. This permits a robust mixed-precision support system across various formats, including FP64, FP32, FP16, and INT8, catering to a vast array of AI model requirements.

Early tests on the system’s CPU throughput using the DeepSeek benchmark have shown promising results, achieving a rate of 578 Tokens Per Second. The overall throughput is projected to be 100 times greater. In addition to compatibility with mainstream models like Qwen, the LineShine supercomputer will be equipped to handle various domestic AI applications and serve numerous domains including remote sensing, materials science, bioinformatics, meteorology, pharmaceuticals, oil exploration, life sciences, and electromagnetic simulations.

Though a specific operational schedule has yet to be established, industry experts predict that the LineShine supercomputer may be operational by around 2029-2030, aligning with China’s growing technological demands.

News Source: OGAWA, Tadashi

Source&Images