China's LineShine Supercomputer Tops TOP500 with 2.2 Exaflops
At ISC 2026 in Hamburg, the 67th TOP500 list revealed a new #1: the LineShine supercomputer in Shenzhen, China. It's the first Chinese submission in nine years, and it's a CPU-only behemoth.
The LX2 CPU: 304 Cores, 228 MB L2 Cache
LineShine is powered by the LX2, an Armv9-compliant CPU with SVE2 and SME support. Each core has 32 KB L1 instruction and 32 KB L1 data cache. The chip consists of two compute dies, each containing four 40-core clusters. Two cores per cluster are disabled, leaving 38 active cores per cluster—152 per die. Each cluster has 28.5 MB of L2 cache, giving each die 114 MB of L2, and the full LX2 package 304 active cores with 228 MB of total L2 cache.
The 304 cores run at 1.55 GHz, delivering 60.3 TFLOP/s FP64 at 690 watts per LX2. The package includes eight stacks of "high-bandwidth memory," 4 GB each (32 GB total) with 4 TB/s bandwidth. This memory might not be conventional HBM—it could be an indigenous Chinese development. Because 32 GB is small for a CPU of this scale, each LX2 is also backed by 256 GB of DDR5 as a spillover tier.
From Node to System: 13 Million Cores
Each node has two LX2 CPUs with 800 Gbps networking (1.6 Tbps per node). Eight nodes form a compute blade, 16 blades make a compute frame, and two frames fill a compute cabinet. The full system has 90 compute cabinets, totaling over 22,000 nodes and 13 million CPU cores.
Benchmarks: #1 in LINPACK and HPCG
LineShine achieved 2.198 exaflops sustained FP64 (Rmax) out of 2.735 exaflops theoretical (Rpeak). It consumed 42.22 MW, yielding 52.07 Gigaflops per watt FP64 efficiency—well behind the Green500 leader (73.282) but impressive for a CPU-only system. Unlike prior Chinese supercomputers, LineShine is not a "LINPACK-special": it also leads the HPCG benchmark with 22.004 petaflops, beating El Capitan's 17.406 petaflops.
Other Notable Systems
- HPC7 (Eni, Italy): New #6, functionally a 30%-scale El Capitan. Uses HPE Cray EX4000 with AMD Instinct MI300A APUs. Rmax 571.5 petaflops FP64 at 8.735 MW. Italy now has more TOP500 compute than any other European country.
- Fugaku: Former #1, now #9, but still #3 on HPCG—a testament to its HPC-focused design despite being six years old.
Green500: No Changes in Top 10
For the first time, the Green500 Top 10 remained unchanged. However, overall HPC efficiency increased due to retirements of older systems.
Implications and Questions
The LineShine submission raises questions: Will China submit results for its other exascale systems (Sunway Oceanlight, CNIS)? Will the US DOE increase funding for large systems? Eni, an oil and gas company, now has two of the Top 10 systems. Why don't large AI clusters (e.g., xAI's Colossus 2 with 550,000 Blackwell GPUs) submit to TOP500? A full HPL run is an excellent stress test for compute, memory, and networking.
Finally, the ISC Group will hand over TOP500 management to ACM SIGHPC, meaning future lists will have DOI numbers for easier referencing.
What Developers Should Know
- Armv9 with SVE2/SME: The LX2 is a real-world, large-scale implementation of these instruction sets. Expect more HPC software to optimize for them.
- CPU-only exascale is possible: LineShine proves that CPUs can reach exascale without accelerators, though at lower efficiency than hybrid systems.
- HPCG matters: LineShine's HPCG lead shows that real-world scientific workloads (sparse linear algebra) can be optimized on CPU-only architectures.
- China's HPC ecosystem is maturing: Indigenous memory technology and massive Arm-based designs signal a shift in the global HPC landscape.
