[Zheshang AI Team] # The Underestimated King of General-Purpose Computing in the AI Era → CPU

English summary

In the training phase, GPU is king (typical configuration CPU:GPU=1:12), but models need to land and monetize; inference cloud services are the solution carrier—#at this time CPU's general-purpose computing power is irreplaceable: business orchestration, virtualization, thousands of industries going to cloud, CPU general computing importance highlighted. Four signals verify CPU status recovery: 1. Supernode ratio flip — Huawei 384 supernode CPU:GPU is 1:2, 8192 supernode CPU/GPU more flexible; Intel/AMD believe future 1:1 or even higher. 2. Unknown algorithm exploration — robot control, embodied intelligence, physical AI, CPU versatility solves the algorithm versatility problem. 3. Inference cost better — ARM adds AI instruction set to CPU, RISC-V explores CPU inference solutions, lower cost. 4. ...