AI Server Evolution and Parallelism

Servers have gone through an evolution of four models with scenario requirements: general purpose servers, cloud servers, edge servers, and AI servers.AI servers have enhanced their parallel computing capabilities by adopting GPUs to better support the needs of AI applications;
AI servers can be divided into two types, training and inference, based on application scenarios. The training process requires high chip arithmetic, and according to IDC, the proportion of inference arithmetic demand is expected to rise to 60.8% by 2025, with the widespread application of large models;
AI servers can be divided into CPU+GPU, CPU+FPGA, CPU+ASIC and other forms of combination according to the type of chip. At present, the main choice in China is the CPU+GPU combination, accounting for 91.9%;
The cost of AI servers comes mainly from chips such as CPUs and GPUs, which take up anywhere from 25 to 70 per cent of the total cost. For training servers, more than 80 per cent of the cost comes from the CPU and GPU.