The problem lies in the architecture. LineShine is a colossal cluster built exclusively on central processing units (CPU-only). Such a topology is ideal for classic HPC (High-Performance Computing) tasks: fluid dynamics, quantum chemistry, or climate modeling. But training modern LLMs requires matrix tensor calculations, which only GPUs can handle efficiently. This release effectively proves the success of US sanctions: cut off from advanced NVIDIA and AMD accelerators, China is forced to extensively build up CPU capacities, which are physically incapable of delivering a breakthrough in frontier machine learning models.
Source: Reuters / TOP500
HardwareGeopoliticsChinaHPCCompute