Each discussion on large-scale artificial intelligence invariably leads to the ubiquitous presence of GPUs in data centers. These chips are engineered for parallel processing, which makes them ideal for training AI models and executing them at scale. Typically, we assume that more AI implies a greater reliance on GPUs. However, China is currently experimenting with a different paradigm that focuses on utilizing CPUs exclusively for its AI infrastructure.<\/p>\n
China has initiated the deployment of several CPU-only supercomputers for high-performance computing and AI workloads due to restrictions imposed by the U.S., curbing access to cutting-edge GPUs. This choice stems not just from technical preferences but is significantly influenced by geopolitical considerations. When access to advanced hardware is restricted, the logical recourse is to innovate via domestic architecture, thereby mitigating dependence on external resources.<\/p>\n
A notable example of China\u2019s CPU-centric approach is the LineShine supercomputer, associated with the National Supercomputing Center in Shenzhen. This machine is entirely constructed with homegrown CPUs, designed specifically to function without any GPUs. As reported by the South China Morning Post<\/strong>, Huang Xiaohui, the center\u2019s deputy director, has emphasized the system’s architecture, which can efficiently handle traditional high-performance computations alongside AI tasks. The configuration includes 47,000 CPUs distributed across 92 cabinets.<\/p>\n At the heart of this initiative lies the LX2 processor, an Armv9 chip explicitly tailored for AI and high-performance computing. Each CPU houses two chiplets and contains 304 cores, which are systematically organized into eight clusters of 38 cores each. This architecture integrates Arm SVE and SME units designed to accelerate vector and matrix operations, crucial for AI training and computational tasks. Furthermore, it uses a combination of High Bandwidth Memory (HBM) and external DDR5 to ensure efficient data transfer, optimizing performance without compromising on capacity.<\/p>\n LineShine aims to achieve groundbreaking performance levels, targeting 2 exaflops\u2014an ambitious goal positioned to surpass the current leader, El Capitan, a supercomputer at Lawrence Livermore National Laboratory, which operates at nearly 1.8 exaflops. During a conference on April 24, Huang stated that the full deployment and activation of LineShine would be complete by the end of 2025, achieving sustained performance exceeding 2 exaflops.<\/p>\n While a CPU-centric machine has its merits for specific applications, it does not negate the advantages GPUs hold in artificial intelligence. For tasks demanding intensive parallelization, GPUs typically outperform CPU-only systems in terms of efficiency and energy consumption. Thus, the trend remains largely in favor of blended architectures that leverage both CPUs for general tasks and GPUs for high-performance computing. LineShine showcases an alternative strategy that may serve specific use cases rather than signaling the end of GPU dominance.<\/p>\n In conclusion, China’s pivot to CPU-only supercomputers marks a significant shift in its high-performance computing strategy, catalyzed by external constraints. While this presents promising advancements, the challenges posed by the supremacy of GPUs in AI applications remain a pivotal consideration for the industry.<\/p>\nThe Architecture Behind LineShine: LX2 Chips<\/h3>\n
The Ambitious Performance Goals<\/h3>\n
Challenges and Limitations<\/h3>\n