TOP500 Names New Number One at ISC'26

China's return to the TOP500 top spot after a nine-year absence is significant for national HPC capability and geopolitics -- but it is not, directly, a signal about who leads in AI training or inference. LineShine is a CPU-only exascale system. For transformer training, inference, or fine-tuning workloads, GPU-dense clusters remain the relevant metric, and those are overwhelmingly built on NVIDIA hardware. The benchmark China just won (High Performance Linpack, a dense linear algebra test) and the benchmark that matters for AI capacity (GPU FLOPS, memory bandwidth, interconnect) measure different things.

LineShine, hosted at the National Supercomputing Centre in Shenzhen (NSCS), delivered 2.198 Exaflops sustained Rmax -- more than 20% ahead of former #1 El Capitan -- across roughly 22,000 nodes with 13,789,440 CPU cores total. The processor is the LX2: an Armv9-compliant chip with SVE2 and SME support, 304 active cores per package running at 1.55 GHz, rated at 60.3 TFLOP/s FP64 at 690 watts. Each node pairs two LX2s with 1.6 Tbps of networking. The system draws 42.22 MW at 52.07 GFlops/W -- well behind the NVIDIA-led Green500 efficiency leaders, but notable for a CPU-only design. One detail worth tracking: the LX2 uses what appears to be an indigenous Chinese on-package memory technology, described in HBM-like terms but likely not conventional HBM. If accurate, that signals Chinese progress in memory technology even under export restrictions.

Unlike some prior Chinese supercomputer submissions that were suspected of being benchmark-optimized, LineShine also tops the HPCG ranking (22.004 Petaflops vs El Capitan's 17.406), which stresses memory bandwidth and sparse compute -- patterns closer to real scientific workloads. This is the first time a CPU-only system has sustained over two Exaflops on the standard HPL benchmark. Fugaku, the previous #1, sits at #9 on TOP500 but retains #3 on HPCG after six years, a testament to its memory-bandwidth-focused design.

Despite LineShine taking the HPL crown, NVIDIA's share of AI-oriented supercomputing tightened. NVIDIA powers 81% of all TOP500 systems and 90% of systems new to this edition. GPU adoption hit a record 238 systems; NVIDIA networking a record 376 systems. The Green500 top 8 -- which measures energy efficiency, a key constraint for large AI deployments -- are all NVIDIA GPU systems. The leader, KAIROS at the University of Toulouse in France, delivers 73.3 GFlops/W on a Grace Hopper Superchip. Italy's new HPC7 (a scaled-down El Capitan using HPE Cray EX4000 and AMD Instinct MI300A APUs) enters at #6; Italy now has more compute on the TOP500 than any other European country.

First, China is known to have at least two other exascale systems -- Sunway Oceanlight and CNIS -- that have not been submitted to the TOP500. Whether they appear on the November list will signal how much China is willing to disclose. Second, large AI training clusters like xAI's Colossus 2 (reported at over 550,000 Blackwell GPUs) have also not been submitted. The gap between TOP500 rankings and actual AI training infrastructure is widening.

The ISC Group is transferring the TOP500 list to ACM SIGHPC. Going forward each edition will carry a dedicated DOI, making citations more reproducible -- a meaningful change for researchers who use TOP500 data as benchmarking baselines.