Model training in a heterogeneous environment

This is not a very general case. The motivation for heterogeneous DNN model training is that we may have some legacy hardware accelerators. For example, a company may have used NVIDIA K80 GPUs 10 years ago. Now the company purchases new GPUs such as NVIDIA V100. However, the older K80 GPUs are still usable and the company wants to use all the legacy hardware.

One key challenge of doing heterogeneous DNN model training is load balancing among different hardware.

Let's assume the computation power of each K80 is half of the V100. To achieve good load balancing, if we are doing data parallel training, we should assign N as the mini-batch size on K80 and 2*N as the mini-batch size on V100. If we are doing model-parallel training, we should assign 1/3 layers on K80 and 2/3 layers on V100.

Note that the preceding example for heterogeneous DNN training is simplified. In reality, it is much harder to achieve decent load balancing...