TL;DR
China has deployed the ‘LineShine’ supercomputer, a CPU-only system delivering 1.54 exaflops, to bypass US-imposed GPU bans. It highlights China’s focus on domestic supercomputing capabilities and AI workloads.
China has deployed the ‘LineShine’ supercomputer, a CPU-only system that achieves 1.54 exaflops of performance, bypassing US restrictions on GPU exports. This development underscores China’s efforts to develop independent supercomputing infrastructure for AI and scientific research.
The ‘LineShine’ supercomputer, operated by China’s National Supercomputing Center, comprises 20,480 nodes, each with two LX2 processors, totaling over 2.45 million CPU cores. These processors are based on Armv9 architecture, featuring 304 cores per chip and integrated with high-bandwidth memory subsystems, including 32 GB of HBM and 256 GB of DDR5 memory.
Each LX2 processor delivers 60.3 TFLOPS FP64 and 240 TFLOPS BF16/FP16, with peak INT8 performance reaching 960 TOPS. The entire system achieves 1.54 exaflops in BF16 training performance, peaking at 2.16 exaflops during specific AI training tasks. The supercomputer is interconnected via a high-speed LingQi network, providing 1.6 Tb/s per node.
Unlike typical supercomputers relying on GPUs for AI workloads, ‘LineShine’ uses only CPUs, which allows for unified memory access, simplified programming, and integration with traditional HPC tasks. This approach is partly driven by US export restrictions on advanced GPUs, prompting China to develop self-reliant AI compute infrastructure.
Why It Matters
This development marks a strategic shift in China’s supercomputing landscape, emphasizing CPU-based architectures to circumvent US export bans on high-performance GPUs. It demonstrates China’s focus on building independent AI and HPC capabilities, potentially reducing reliance on foreign technology and reshaping global supercomputing dynamics.
The system’s high performance in AI training tasks also suggests a move towards CPU-centric solutions for certain scientific and industrial applications, challenging the industry norm of GPU dominance in exascale computing.
CPU Ryzen® 7 9800X3D Solutions Express Gaming PC | RTX 5070 w/12GB | 8 Core 5.2Ghz | 64GB RAM | 2000GB NVMe PCIe SSD | Black RGB Tower |
Experience the unparalleled power of the AMD Ryzen 7 9800X3D Processor. This high-performance CPU features an impressive 8…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
Over recent years, US restrictions have limited China’s access to high-end GPUs, prompting a pivot towards CPU-only supercomputers for AI and HPC workloads. China’s National Supercomputing Center previously deployed CPU-centric systems, but ‘LineShine’ represents a significant leap with its exaflop-scale performance using Arm-based processors with advanced memory subsystems.
Most leading supercomputers globally rely on heterogeneous architectures combining CPUs and GPUs, especially for AI. China’s move to CPU-only systems indicates a strategic effort to develop self-sufficient infrastructure and reduce dependence on US technology amid ongoing geopolitical tensions.
“The ‘LineShine’ supercomputer exemplifies China’s shift towards CPU-centric AI and HPC systems, aiming to bypass US export restrictions on GPUs while achieving exascale performance.”
— Anton Shilov, Tom’s Hardware
“The system is designed to support both AI training and traditional HPC workloads, with a focus on independence and high performance.”
— China’s National Supercomputing Center
Armv9 server processors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Remains Unclear
It remains unclear how the ‘LineShine’ system’s real-world efficiency compares to GPU-based supercomputers in various applications, as peak theoretical performance may differ from operational throughput. Details about its power consumption and long-term stability are also still emerging.
ADATA DDR5 5600 SO-DIMM Memory Module – 16GB High Bandwidth Laptop Memory Module (RAM) – High-Speed 5600MHz – Automatic Error Correction – Compatible with AMD & Intel Platforms – AD5S560016G-S
Compatible for select DDR5 Laptop, Notebook, Mini PC, and All-in-One (AIO) Computers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What’s Next
Further performance benchmarking and integration tests are expected in the coming months. Monitoring how ‘LineShine’ performs in real AI and scientific tasks will clarify its role relative to GPU-based systems. Additionally, China may develop more CPU-centric supercomputers or attempt to optimize existing architectures for broader applications.
ST-JY PCIe 4.0 x4 Oculink SFF-8611 4i to SFF-8611 4i High-Speed Data Cable, 64Gbps Bandwidth for AI GPU, Servers, Data Center, External Storage/Graphics Expansion (80cm)
Supports PCIe 4.0 protocol, delivering a total bandwidth of up to 64 Gbps (~8 GB/s). Unleashes the full…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Why did China develop a CPU-only supercomputer?
Due to US export restrictions on GPUs, China aimed to create a self-reliant supercomputing infrastructure capable of high-performance AI and HPC workloads without relying on foreign GPU technology.
How does ‘LineShine’ compare to GPU-based supercomputers?
While ‘LineShine’ achieves 1.54 exaflops in BF16 training, its peak theoretical FP64 performance is around 2.47 exaflops. Compared to GPU-based systems, it may be less power-efficient but offers advantages in integration, memory coherence, and reduced dependency on foreign hardware.
What are the main advantages of CPU-only supercomputers?
They simplify programming, avoid costly data transfers between CPU and GPU, support larger coherent memory pools, and integrate more seamlessly with traditional HPC workflows.
Will China replace all GPU supercomputers with CPU-only systems?
It is unlikely in the near term. While CPU-only systems like ‘LineShine’ are significant for independence and specific applications, heterogeneous CPU-GPU architectures remain dominant for many high-density AI tasks due to efficiency advantages.
