China Fields 1.54-Exaflops LineShine Supercomputer Built Entirely on Armv9 CPUs

*China’s National Supercomputing Center in Shenzhen has deployed a 1.54-exaflops system that uses only Huawei-designed Armv9 processors, avoiding any reliance on restricted GPUs.*

The build

The machine contains 2.4 million LineShine LX2 cores. Each core follows the Armv9 architecture developed by Huawei. The design draws directly from the approach taken by Japan’s Fugaku supercomputer and its A64FX processor, substituting a domestic CPU for the GPU clusters that have dominated recent exascale systems.

No graphics processors appear in the reported configuration. All floating-point work runs on the general-purpose Armv9 cores. The result places the system at 1.54 exaflops while remaining fully compliant with current US export controls on advanced accelerators.

Technical choices

Shenzhen engineers scaled the design by increasing core count rather than clock speed or specialized matrix units. The same strategy allowed Fugaku to reach the top of the performance charts in 2020 without GPUs. By following that template with Huawei silicon, the center achieved comparable throughput on workloads that can be expressed in standard Armv9 instructions.

Power and cooling requirements differ from GPU-heavy installations. A CPU-only layout spreads heat across more, simpler dies and reduces the need for high-bandwidth memory stacks that often accompany accelerator cards. These differences matter when domestic supply chains must substitute for embargoed components.

Limits of the data

Public details stop at peak floating-point throughput and core count. No sustained performance figures on standard benchmarks such as HPL or HPCG have been released. It is therefore unclear how the machine compares with GPU-based systems once real applications, memory bandwidth, and interconnect limits are taken into account.

Why it matters

US export rules have treated GPUs as the decisive bottleneck for frontier computing. LineShine shows that an alternative path remains open: accept lower per-core efficiency and compensate with volume. For organizations that must operate under the same restrictions, the lesson is practical rather than theoretical. Large-scale Armv9 deployments can still deliver headline exaflops numbers. Procurement plans that assume GPUs are indispensable now carry an observable counter-example.

---

Sources:

• Tom's Hardware

{
  "excerpt": "China’s Shenzhen supercomputing center has built a 1.54-exaflops, CPU-only machine using 2.4 million Huawei Armv9 cores to sidestep US GPU export controls.",
  "suggestedSection": "hardware",
  "suggestedTags": ["supercomputing", "china", "arm", "huawei"],
  "imagePrompt": "Rows of identical metallic compute racks stretch into a dim industrial hall, their uniform surfaces broken only by faint status lights along each cabinet edge. The scene emphasizes scale and repetition rather than individual machines. muted color palette, cinematic lighting, 16:9"
}