China's LineShine tops Top500 with 2.2 ExaFLOPS

Key Takeaways

• LineShine hit 2.198 ExaFLOPS in the Linpack benchmark, 26% faster than El Capitan's 1.742 ExaFLOPS
• The Chinese system is the first CPU-only machine to break the 2 ExaFLOPS barrier
• El Capitan still leads in energy efficiency (60.94 vs 52.07 GFLOPS/W) and mixed-precision AI workloads

China's LineShine supercomputer has taken the top spot on the Top500 list, dethroning the US's El Capitan with a Linpack benchmark score of 2.198 ExaFLOPS. The machine, deployed at the National Supercomputing Centre in Shenzhen, is the first in the ranking's history to break 2 ExaFLOPS using only CPUs.

This marks China's return to the top of supercomputing rankings for the first time since 2019. El Capitan, located at Lawrence Livermore National Laboratory and powered by AMD's MI300A accelerators, held the crown since November 2024 with 1.742 ExaFLOPS. LineShine beats it by roughly 26%.

What powers the LineShine supercomputer?

The system runs on semi-custom 304-core LX2 processors based on Arm's v9 instruction set, clocked at 1.55 GHz. Each chip uses two compute chiplets with eight clusters of 38 cores apiece. The total core count across the machine: 13.79 million.

Memory architecture is unusual. Each LX2 pairs 32 GB of on-package HBM (offering up to 4 TB/s bandwidth) with up to 256 GB of external DDR5. The idea is to maximize both bandwidth for intensive calculations and capacity for large datasets.

The LX2 includes Arm's SVE (Scalable Vector Extension) and SME (Scalable Matrix Extension) units for accelerating vector and matrix operations. Supported data formats span FP64, FP32, BF16, FP16, and INT8. The system connects via a proprietary interconnect called LingQi.

Where El Capitan still wins

Raw performance is not the whole picture. LineShine consumes 42.2 MW of power, delivering 52.07 GFLOPS per watt. El Capitan manages 60.94 GFLOPS/W. For organizations paying electricity bills at scale, that 17% efficiency gap matters.

More significant: LineShine trails badly on mixed-precision AI workloads. In the HPL-MxP benchmark (which tests low-precision performance relevant to machine learning training and inference), LineShine scored 7.92 mixed-precision EFLOPS. El Capitan, Frontier, and Aurora all beat it. The LX2's uplift when moving from FP64 to mixed-precision is only 3.6x, far lower than GPU-accelerated systems with dedicated low-precision hardware.

Tom's Hardware notes this limits LineShine's usefulness for AI training, though it excels at traditional scientific computing tasks.

Why China submitted these results

China has operated world-class supercomputers for years but stopped submitting to Top500 after 2019, likely to avoid US scrutiny of its chip supply chains. That NSCS submitted LineShine's results now signals confidence. The organization apparently believes the machine relies entirely on domestic technologies that US export controls cannot touch.

This is a direct response to years of US sanctions aimed at cutting China off from advanced semiconductors. Whether the LX2 processors are fully produced within China (including fabrication) remains unclear from available information. But the submission itself is a statement.

How does LineShine compare to other CPU-only systems?

The obvious comparison is Japan's Fugaku, which held the top spot from 2020 to 2022 using only Arm-based A64FX CPUs. Fugaku manages 14.78 to 16.84 GFLOPS/W depending on optimization. LineShine triples that efficiency while delivering roughly five times Fugaku's peak performance.

What this means for supercomputing competition

The US still operates three of the top five machines: El Capitan, Frontier, and Aurora. Europe's JUPITER ranks sixth. The race is not about one machine. It is about sustained investment, domestic chip capability, and the ability to build at scale without foreign dependencies.

LineShine proves CPU-only designs can reach exascale without GPUs. That matters because GPU supply is constrained and expensive. Whether this architecture makes sense for workloads beyond traditional HPC, particularly AI, remains the open question.

Frequently Asked Questions

What makes LineShine different from other supercomputers?

It is the first system to exceed 2 ExaFLOPS using only CPUs, without GPU accelerators. Most exascale machines rely heavily on GPUs for peak performance.

Is LineShine better than El Capitan?

LineShine is 26% faster on the Linpack benchmark. But El Capitan is more energy-efficient (60.94 vs 52.07 GFLOPS/W) and significantly better at mixed-precision AI workloads.

Why did China submit to Top500 after years of absence?

The submission signals confidence that LineShine uses entirely domestic technologies unaffected by US export restrictions.

What processor does LineShine use?

Custom 304-core LX2 processors based on Arm's v9 architecture, running at 1.55 GHz with on-package HBM and DDR5 memory.

Can LineShine train AI models effectively?

Its mixed-precision performance lags behind GPU-accelerated systems like El Capitan. LineShine is optimized for traditional scientific computing, not AI training.

Source: Latest from Tom's Hardware