of china LineShine supercomputerbecame the fastest system in the world, surpassing El Capitan in the June 2026 TOP500 list. System in HPL Linpack test 2,198 exaFLOPS FP64 performance and became the first supercomputer to exceed the 2 exaFLOPS threshold among CPU-based machines. Moreover, this result showed that China has made a strong return to the TOP500 list after a three-year break. Working in Shenzhen, LineShine stood out this time with domestic processors, not GPU accelerators.
According to TOP500 data, LineShine operates at the National Supercomputing Center in Shenzhen. Shenzhen Cloud Computing Center built the system on the LingKun platform and 304-core LX2 processors used. However, the machine is powered by Armv9-based semi-custom processors running at a frequency of 1.55 GHz. Also in the system 13,789,440 coresLingQi interconnect and Kylin OS operating system.
Power consumption of LineShine 42.2MW reaches the level. In contrast, the system per watt 52.07 GFLOPS offers efficiency and in this area El Capitan 60.94 GFLOPS/W falls short of its results. On the other hand, LineShine clearly surpasses Fugaku, one of the old CPU-focused leaders. Fugaku, according to optimization preference 14.78 to 16.84 GFLOPS/W It was working with efficiency values varying between .
El Capitan uses AMD EPYC processors with AMD Instinct MI300A accelerators. LineShine, on the other hand, leans entirely on the CPU side and therefore the comparison has a different meaning. Additionally, since the TOP500 list is centered on the HPL result, LineShine’s FP64-focused scientific computing performance played a decisive role in the ranking. In addition, the system is also on the HPCG list. 22.00 HPCG-PFLOPS He ranked first with his result.
LineShine changed the TOP500 ranking with CPU power
On the mixed sensitivity side, the situation is different. LineShine in HPL-MxP test 7.92 exaFLOPS reached its conclusion and El Capitan fell behind Frontier and Aurora. This result means that the system artificial intelligence training and inference It shows that it is not as high as its GPU-accelerated competitors in its work. Despite this, HPL and HPCG results put the machine at the top of traditional supercomputer workloads.
LX2 processor, a total of two compute chipsets 304 CPU cores It carries. There are eight CPU clusters with 38 cores per processor, and each core leverages Arm SVE and SME units. Additionally, these units accelerate vector and matrix operations in FP64, FP32, BF16, FP16 and INT8 data formats. This technical foundation enables the processor to be expanded to a wider task area in scientific computing and certain AI workloads.
On the memory side, the LX2 follows a different path from conventional CPU designs. Processor, 32 GB in-package HBM memory 4 TB/s uses bandwidth and Up to 256 GB external DDR5 memory is accompanied. Thus, the system gathers calculation tasks requiring high bandwidth and scenarios requiring large memory capacity around the same processor. However, this architectural choice does not fully capture the leap forward offered by dedicated accelerators in low-precision calculations.
LineShine when switching from FP64 to mixed precision data 3.6 times provides performance increase. Systems using accelerators such as AMD Instinct MI300A or Intel Ponte Vecchio go to higher multipliers for low-precision tasks. Additionally, memory bandwidth, software maturity and interconnect efficiency play a role in this difference. Therefore, LineShine stands out in the FP64-centric supercomputer performance rather than in the AI race.
Sending results to the TOP500 also carries a different message. The National Supercomputer Center in Shenzhen included LineShine data in the list, making the technical level of the system visible. Additionally, the evaluation in the source text conveys that the center relies on domestic technologies at LineShine and thinks that US restrictions will not directly affect this production process. Thus, China rose to the top of the TOP500 rankings again and opened a new page in the field of high-performance computing.