By China’s LineShine just took the top spot on the TOP500 supercomputer rankings — and it got there without a single GPU. Announced at ISC 2026 in Hamburg on June 23, the machine hit 2.198 ExaFLOPS on the HPL benchmark, landing roughly 21 percent ahead of El Capitan. It is China’s first #1 on TOP500 since Sunway TaihuLight in 2017. What makes this genuinely interesting isn’t the raw number. It’s the architecture.
13.79 Million ARM Cores, No GPUs Required
LineShine runs on Huawei-designed LX2 processors — an ARMv9 chip with 304 cores per socket, running at 1.55 GHz. Spread across 20,480 nodes, the system reaches 13.79 million total compute cores. There is no NVIDIA H100. No AMD Instinct. No Intel Gaudi. The entire stack — processors (LX2/LingKun), interconnect (LingQi), operating system (Kylin), and storage — is Chinese-designed and domestically produced.
The LX2 isn’t just a server CPU grafted onto a big cluster. Each chip includes ARM Scalable Vector Extension (SVE) and Scalable Matrix Extension (SME) units that handle FP64, BF16, and INT8 workloads natively. Per-chip throughput is 60.3 TeraFLOPS at FP64, 240 TeraFLOPS at BF16, and 960 TeraOPS at INT8. At 13.79 million cores, that scales.
Where It Leads — and Where It Doesn’t
LineShine wins two of the three major TOP500 benchmarks. On HPL (standard double-precision Linpack), it posts 2.198 ExaFLOPS to El Capitan’s 1.809. On HPCG, which better reflects real-world scientific workloads like finite element simulation and sparse algebra, LineShine takes first at 22.00 PetaFLOPS versus El Capitan’s 17.41.
| Benchmark | LineShine | El Capitan |
|---|---|---|
| HPL (FP64) | 2.198 ExaFLOPS (#1) | 1.809 ExaFLOPS (#2) |
| HPCG (real-world) | 22.00 PF (#1) | 17.41 PF (#2) |
| HPL-MxP (AI) | 7.92 ExaFLOPS (#4) | 16.7 ExaFLOPS (#1) |
| GPU accelerators | None | AMD MI300A |
| Power draw | 42.2 MW | ~29 MW |
Then there is HPL-MxP, the mixed-precision benchmark designed to capture AI-relevant workloads. Here the picture shifts. El Capitan scores 16.7 ExaFLOPS on HPL-MxP — a 9.2x multiplier over its standard HPL score, which is what GPU hardware does with low-precision matrix math. LineShine scores 7.92 ExaFLOPS, a 3.6x multiplier. That gap is not a rounding error. It is the direct consequence of having no dedicated tensor cores or GPU-style low-precision accelerators.
For AI training at transformer scale, GPU-accelerated clusters still dominate. LineShine is not a GPU replacement for building frontier models. Call it what it is: the world’s fastest traditional HPC machine that also does reasonable BF16 throughput, but not in the league of GPU-scale AI compute.
Export Controls Met Their Match — Sort Of
The subtext here is impossible to ignore. Since 2022, the US has progressively restricted China’s access to advanced AI chips: A100, H100, H200, and equivalent AMD parts are all blocked. The stated goal was to slow Chinese AI development by constraining compute. LineShine is China’s answer, and it is a credible one in the HPC domain.
SMIC, China’s main chip fab, remains constrained around 7nm for high-bandwidth memory and advanced logic. But Huawei’s LX2 processors are running in production at HPC scale, and the entire software stack is domestic. Chinese officials described it as “complete self-reliance and controllability.” That is not hyperbole for this system.
Whether export controls slowed China’s AI trajectory remains genuinely debated. What LineShine proves is that they did not prevent China from building the world’s fastest supercomputer on its own terms.
What Developers Should Actually Take From This
If you are not running climate simulations or nuclear weapons modeling, LineShine does not change your week. But a few things are worth tracking.
ARM at exascale is now proven. AWS Graviton, Microsoft Cobalt, and Ampere Altra are all descendants of the same ARM architecture that now powers the world’s #1 supercomputer. Cloud providers have strong incentive to push CPU-only inference pipelines harder. If your workload runs well on BF16/INT8 without needing gradient backpropagation, a dense ARM cluster may be a serious option within the next few years.
Sovereign compute is accelerating. Governments and enterprises unable or unwilling to depend on US-aligned GPU supply chains now have proof of concept. Expect more investment in CPU-optimized AI software — MLIR, IREE, and compiler toolchains targeting non-NVIDIA hardware will get more funding and more production usage.
And finally: the compute geography is fragmenting. US GPU clusters on one side, China CPU HPC on another, European and Asian sovereign compute in between. Developers building global infrastructure or AI products will increasingly need to design for heterogeneous compute access.
LineShine is a genuine milestone. The nuance is in which milestones it actually represents — and which ones it doesn’t.
