Being an ML architect

Interview with Yangqing Jia: https://developer.aliyun.com/article/742724

Nowadays, with cloud service, it is no longer necessary to have someone working to build the GPU cluster. Then, what next?

Each company should have an ML architect. For example, if I bought 1000 GPU cards, what's the utilization rate? what's the usage? Whether it is better to run in GPU or CPU? How to integrate the sub-modules with the models.