Every time we talk about large-scale artificial intelligence we end up reaching the same point: data centers and their enormous amounts of GPUs. It’s not a coincidence. This type of chip has become a centerpiece because it is especially well suited to running many operations in parallel, just the kind of work that requires training AI models and running them at scale. We take it almost for granted: more AI, more GPU. But that equivalence does not exhaust all possibilities. China is trying a different route, one that attempts to answer the same question from another place: what happens if AI muscle is built with CPUs alone.
CPU instead of GPU. HPC Wire notes that China has begun deploying several CPU-only supercomputers in recent years for AI workloads and high-performance computing, largely due to US restrictions that limit its access to enough advanced GPUs for these types of systems. The difference is important: we are not simply talking about a technical preference, but rather a response conditioned by the geopolitical context. When access to the most coveted hardware is limited, the alternative is to squeeze out our own architectures and reduce external dependency.
LineShine. The most striking case of this strategy is this supercomputer, linked to the National Supercomputing Center in Shenzhen. According to the South China Morning Post, it is a machine built entirely with domestic CPUs and designed to run without a GPU. The media also reports that Huang Xiaohui, deputy director of the Shenzhen center, presented it as an integrated architecture capable of supporting both traditional high-performance computing and artificial intelligence loads. The system, they explain, uses 47,000 CPUs spread across 92 computing cabinets.
The LX2 chip. The piece that allows us to lower that bet into the realm of hardware is the LX2 processor, described as an Armv9 chip designed for AI loads and high-performance computing. Each CPU integrates two chiplets and has 304 cores, organized in eight clusters of 38 cores each. The architecture includes Arm SVE and SME units, designed to accelerate vector and matrix operations, widely present in AI training and scientific computing. Added to that is an unusual combination of HBM memory in the package itself and external DDR5, a mix aimed at moving a lot of data quickly without giving up capacity.
The power. LineShine is designed to reach 2 exaflops, a figure with which China aspires to place it above El Capitan, the Lawrence Livermore National Laboratory supercomputer that is the current world leader with almost 1.8 exaflops. Huang Xiaohui, deputy director of the Shenzhen center, went further at a conference on April 24. According to statements collected by SCMP, he maintained that by the end of 2025 the system had completed its deployment and activation, with sustained performance greater than 2 exaflops.
Not everything is positive. Going for a CPU-only machine may make sense for certain jobs, but it doesn’t eliminate the great advantage of GPUs in artificial intelligence. For more intensive and easily parallelized loads, these accelerators typically complete more work with the same power than a CPU-only system. That is why the industry continues to rely mostly on mixed architectures, with processors for general tasks and GPUs to accelerate heavier calculations. LineShine fits better as an alternative route under specific conditions than as proof that the dominant model is behind us.
Images | with Nano Banana
In | There was a time when Nvidia was a gaming company. That business is now pocket change for the owner and lady of AI
