China has reclaimed the top spot on the TOP500 list of the world's fastest supercomputers for the first time in nearly a decade. The system, called LineShine and located in Shenzhen, debuted at number one on the 67th edition of the ranking, which was unveiled at the ISC 2026 conference in Hamburg, Germany.
The result is a milestone not just for raw performance but for what it signals about China's self-sufficiency in high-performance computing. For the latest AI hardware developments and supercomputing coverage, the LineShine entry marks China's first formal TOP500 submission in nine years.
A CPU-Only Giant
What makes LineShine unusual is its architecture. The system is entirely CPU-based — no GPUs. It is built around a custom processor called the LX2, an Armv9-compliant chip with support for SVE2 and SME extensions. Each LX2 package packs 304 active cores running at 1.55 GHz, delivering a quoted 60.3 teraflops of FP64 compute at 690 watts per socket.
The chip is constructed from two compute dies, each containing four 40-core clusters. Two cores per cluster are disabled for yield, leaving 38 active cores per cluster and 152 per die. Each die carries 114 MB of L2 cache, bringing the full package to 228 MB of L2.
Memory is where the design gets particularly interesting. The LX2 includes eight stacks of on-package "high-bandwidth memory" totaling 32 GB with 4 TB/s of bandwidth. Analysts at Chips and Cheese, who reported the benchmark details, noted that this memory is "likely not conventional HBM" and could be an indigenous Chinese development, given that 32 GB is unusually small for a processor of this scale. Each LX2 is also backed by 256 GB of DDR5 memory as a larger spillover tier.
At the node level, each compute node houses two LX2 CPUs paired with 800 Gbps of networking, for a combined 1.6 terabits per second of interconnect bandwidth per node. Eight nodes are grouped into a compute blade, with sixteen blades combining into a rack — a dense, high-fanout topology that explains how a CPU-only design reaches exascale throughput. The reliance on custom Arm-based silicon rather than GPUs is a deliberate divergence from the Nvidia-dominated blueprint that nearly every other top-tier system follows.
The Numbers Behind the Crown
At the system level, LineShine achieved a sustained 2.198 exaflops of FP64 performance (Rmax) on the HPL benchmark, out of a theoretical peak (Rpeak) of 2.735 exaflops. The system drew 42.22 megawatts of power during the run, yielding an efficiency of 52.07 gigaflops per watt — well behind the Green500 leader's 73.28 gigaflops per watt, but described as impressive for a CPU-only machine.
Crucially, LineShine is not merely a benchmark specialist. It also claimed the number-one position on the HPCG benchmark, a more demanding test that better reflects real-world application performance, with a result of 22.004 petaflops per second. That beat the previous HPCG leader, Lawrence Livermore National Laboratory's El Capitan, which scored 17.406 petaflops per second. The Chips and Cheese analysis noted that "unlike prior Chinese supercomputers, this system is not just a LINPACK-special."
China's Long Return
China's absence from the TOP500 has been conspicuous. After the U.S. sanctioned several Chinese supercomputing centers in 2021, Chinese institutions largely stopped submitting results to the list, even as they continued to build systems. The nine-year gap — the analysis dates the last formal submission to roughly 2017 — left the West uncertain about how far China's domestic chipmaking had actually progressed.
The LineShine entry answers that question decisively. The use of a custom Armv9 processor with apparent domestic high-bandwidth memory suggests China has built a viable parallel computing stack that does not rely on Nvidia GPUs or standard HBM from the usual Korean suppliers. For an AI industry that has been forced to find workarounds for U.S. chip export controls, the existence of a competitive domestic HPC platform carries weight well beyond the supercomputing rankings — it demonstrates that sanctioned supply chains can be replaced, at least at the high end of traditional computing.
Other Movers on the List
The 67th TOP500 list included other notable shifts. Eni's HPC7 system in Italy entered at number six, achieving 571.5 petaflops of FP64 performance on the same HPE Cray EX4000 platform and AMD Instinct MI300A APUs used by El Capitan, but at roughly 30% of its scale. With HPC7, Italy now has more compute listed on the TOP500 than any other European country, edging out Germany, which still holds Europe's only other exascale system.
The list also marked an institutional transition. The ISC Group announced it will hand over stewardship of the TOP500 list to ACM SIGHPC, meaning future rankings will carry dedicated DOI numbers, making it easier to reference specific editions of the list in academic work.
For the supercomputing world, the LineShine result is a wake-up call. China has not just returned to the rankings — it has returned at the very top, with homegrown silicon and an architecture that breaks from the GPU-dominated mold.
Stay Ahead of AI
From exascale supercomputers to next-generation AI chips, AI Buzz Wire covers the hardware powering the future of artificial intelligence.
Read more AI news →
