China’s CPU-only supercomputer design represents a fundamental shift in how the world’s second-largest economy approaches frontier computing, abandoning the GPU-accelerator model that dominates the top 500 fastest supercomputers globally. The National Supercomputing Center in Shenzhen is building LineShine, an all-domestic exascale system that relies entirely on Huawei-designed Armv9 processors rather than imported accelerators, a direct response to US export restrictions on advanced semiconductors and AI chips.
Key Takeaways
- LineShine is an all-CPU exascale supercomputer with reported performance of 1.54 exaflops, built entirely on Armv9-based LX2 processors.
- The system contains 20,480 computing nodes, each with two LX2 CPUs delivering 304 cores per processor.
- Each LX2 processor integrates eight on-package HBM stacks with 32 GB capacity and 4 TB/s memory bandwidth.
- Full deployment is expected by the end of 2025, with phase one using 100 Huawei Kunpeng servers totaling 12,800 CPU cores.
- LineShine targets applications including molecular dynamics, fluid simulation, life sciences, and AI model training without reliance on foreign chips.
Why CPU-only supercomputer design matters now
Most exascale systems rely on GPU accelerators from Nvidia or AMD to reach extreme performance levels. LineShine breaks that pattern entirely. By choosing a CPU-only supercomputer design, China demonstrates it can achieve frontier-class computational performance using exclusively domestic silicon, a critical capability given US restrictions on advanced chip exports. This architectural choice echoes Japan’s Fugaku supercomputer and Fujitsu’s A64FX processor, which proved that CPU-heavy systems can compete at the highest performance tiers without sacrificing capability.
The timing is significant. As US export controls tighten around advanced semiconductors and AI accelerators, China faces genuine pressure to develop independent computational infrastructure. LineShine is not just a technological achievement—it is a statement of computational sovereignty. The system’s CPU-only supercomputer design eliminates dependency on foreign accelerators entirely, making it immune to future export restrictions.
LineShine architecture and performance specifications
LineShine’s CPU-only supercomputer design centers on Huawei’s proprietary Armv9-based LX2 processor, a custom chip designed specifically for high-performance computing workloads. Each LX2 integrates two compute dies, delivering 304 cores total per processor. The system scales across 20,480 computing nodes, with each node housing two LX2 CPUs. That architecture yields the reported 1.54 exaflops of performance, though the system is designed to target 2 exaflops as deployment expands.
Memory bandwidth is where LineShine’s CPU-only supercomputer design shows careful engineering. Each LX2 processor includes eight on-package HBM stacks with 32 GB total capacity and 4 TB/s aggregate bandwidth per processor. Off-package, each compute die connects to 128 GB of DDR5 memory across four NUMA domains, creating a memory hierarchy optimized for scientific simulation and AI workloads. The system also features 650 PB of storage with 10 TB/s bandwidth, enabling massive data-intensive applications.
Interconnect is handled by a proprietary high-speed network called LingQi, built entirely from Chinese components. This CPU-only supercomputer design eliminates any reliance on foreign networking technology, completing the domestic-only stack.
Phased deployment and competitive positioning
LineShine is rolling out in phases rather than as a single deployment. Phase one uses 100 Huawei Kunpeng servers with 12,800 CPU cores. Phase two will expand significantly, adding tens of thousands of additional CPUs and scaling the interconnect and storage infrastructure. Full operational deployment is expected by the end of 2025.
How does this CPU-only supercomputer design compare to current leaders? The US Department of Energy’s El Capitan supercomputer at Lawrence Livermore National Laboratory currently tops the TOP500 list with approximately 1.8 exaflops Linpack performance and 2.8 exaflops theoretical peak. El Capitan uses AMD CPUs paired with Nvidia GPUs—the conventional accelerated approach. LineShine’s CPU-only strategy achieves comparable scale without any accelerators, a fundamentally different engineering philosophy. Most top-tier supercomputers follow the GPU-accelerated model because it delivers raw performance per watt more efficiently. LineShine’s architects chose to sacrifice some power efficiency to achieve complete technological independence.
Applications and intended workloads
LineShine is being positioned for large-scale scientific and AI applications: molecular dynamics simulations, computational fluid dynamics, life sciences research, and large-scale AI model training. These workloads benefit from high memory bandwidth and massive core counts—exactly what a CPU-only supercomputer design provides. The system is not optimized for real-time inference or latency-sensitive tasks, where GPU acceleration typically shines. Instead, it targets batch-processing and simulation workloads where sustained throughput matters more than peak performance.
Is China’s CPU-only supercomputer design a viable long-term strategy?
The success of LineShine will determine whether CPU-only supercomputer design can remain competitive as export controls intensify. If the system meets its 2 exaflops target and delivers consistent performance on real-world scientific applications, it proves that domestic CPU design can sustain frontier-class computing without foreign accelerators. If performance falls short or power consumption becomes prohibitive, it may force China to pursue alternative approaches like domestic GPU development. The stakes are high: computational sovereignty in the era of export controls.
What makes LineShine different from other exascale systems?
LineShine is unusual because it pursues CPU-only supercomputer design at exascale. Most contemporary exascale systems use GPU acceleration to reach extreme performance levels. LineShine demonstrates that sheer core count and memory bandwidth can compensate for the absence of specialized accelerators, at least for certain workload classes. This architectural choice reflects geopolitical constraints as much as technical preference.
When will LineShine be fully operational?
Full deployment of the LineShine CPU-only supercomputer design is expected by the end of 2025. Phase one is already underway with 100 Kunpeng servers deployed. Phase two will add the bulk of the system’s computational capacity, scaling to tens of thousands of additional processors and completing the LingQi interconnect infrastructure.
LineShine represents a watershed moment for global supercomputing: proof that a major computational power can build exascale infrastructure entirely from domestic silicon. Whether this CPU-only supercomputer design becomes a model for others or remains a geopolitical necessity unique to China will depend on whether it delivers real-world performance and whether export restrictions continue to tighten. For now, it stands as the clearest signal yet that the era of uncontested US technological dominance in high-performance computing is ending.
Edited by the All Things Geek team.
Source: Tom's Hardware
