The problems of a single CPU+xPU heterogeneous computing itself:
The part that can be accelerated accounts for a limited proportion of the entire system. For example, if the acceleration ratio is 80%, the acceleration will not exceed 5 times;
The impact of data moving back and forth between the CPU and the accelerator reduces the acceleration ratio, and the overall acceleration effect in some scenarios is not obvious;
Heterogeneous acceleration explicitly introduces new entities, and the calculation becomes the explicit cooperation of two or more entities, which increases the complexity of the entire system;
Although the performance of GPU has improved a lot compared to CPU, there is still a significant gap compared with the performance of DSA/ASIC. The problem of DSA/ASIC is that it cannot adapt to the requirements of business flexibility in complex scenarios, resulting in large-scale applications becoming huge threshold;
The CPU+xPU architecture is centered on the CPU. The entire IO path is very long, and IO becomes the bottleneck of performance.