China Aims for Exascale with CPU-Only Supercomputer and 47,000 Domestic Processors

China has announced the Lingshen project, an ambitious initiative aimed at building an exascale supercomputer with a computational capacity of 2 Exaflops. This system stands out due to a peculiar architectural choice: it will be entirely CPU-based, without the use of Graphics Processing Units (GPUs), and will integrate a remarkable 47,000 internally developed processors. The infrastructure will leverage Huawei Kunpeng servers, and a crucial aspect of the project is the complete absence of foreign-made components, underscoring a clear orientation towards technological self-sufficiency.

CPU-Only Architecture: A Strategic Choice

The decision to build an exascale supercomputer based exclusively on CPUs represents a significant deviation from the dominant trend in High-Performance Computing (HPC) and artificial intelligence, which sees GPUs as primary actors for highly parallelizable workloads. While GPUs excel in tasks such as Large Language Model (LLM) training and inference due to their massively parallel architecture, CPUs offer advantages in terms of flexibility, memory management, and the ability to handle a wide range of computational workloads. The deployment of 47,000 homemade processors suggests deep optimization at the silicio and software levels to maximize efficiency and performance with this specific architecture.

Implications for Technological Sovereignty and On-Premise Deployment

The emphasis on the exclusive use of national components, from processors to Huawei Kunpeng servers, highlights China's strategy to strengthen its technological sovereignty. This approach reduces dependence on external supply chains, mitigating geopolitical risks and ensuring complete control over the entire technology stack. For organizations evaluating on-premise deployment of AI/LLM workloads, the Lingshen strategy offers interesting insights. The ability to design and produce hardware internally can translate into unprecedented control over security, customization, and long-term TCO optimization, especially in contexts requiring air-gapped environments or high compliance standards. AI-RADAR provides analytical frameworks at /llm-onpremise to support the evaluation of these trade-offs.

The Future of Supercomputing and AI

The Lingshen project is not only a large-scale engineering endeavor but also an indicator of the strategic directions China intends to pursue in the field of supercomputing and artificial intelligence. While the rest of the world increasingly relies on GPU-based solutions for AI acceleration, Lingshen's CPU-only approach could explore new frontiers in software and hardware optimization for specific workloads. This development underscores the importance of considering diverse architectures and the need to carefully evaluate the trade-offs between performance, cost, energy efficiency, and technological control when planning large-scale AI infrastructures.