Pros and cons of pipeline parallelism
In the preceding sections, we discussed how pipeline parallelism works in both forward and backward propagation during model parallelism training. In this section, we will discuss the advantages and disadvantages of pipeline parallelism.
Advantages of pipeline parallelism
The most important advantage of pipeline parallelism is that it helps to reduce the GPU idle time during model parallelism training. Here, we list all the advantages:
- Reduces overall training time
- Reduces GPU idle time while waiting for the predecessor or successor's GPU output
- Not much coding complexity to implement pipeline parallelism
- Can be generally adapted to any kind of DNN model
- Simple and easy to understand
Disadvantages of pipeline parallelism
In the preceding section, we discussed the advantages of pipeline parallelism. Now let's look at the disadvantages of pipeline parallelism:
- The CPU needs to send more instructions...
