Chapter 8: Achieving Higher Throughput and Lower Latency

Generally speaking, model parallelism is less efficient than data parallelism. The main reasons are twofold, as outlined here.

First, the sequential dependency among deep neural network (DNN) layers holding onto different graphics processing units (GPUs) limits the performance. One GPU may not start working until its predecessor finishes generating outputs.

Second, the limited GPU memory makes it impossible to train a large input batch in each training iteration. Due to the large size of the model parameters, we can only train small batches of data per training iteration.

Given the preceding two challenges, we try to improve throughput and latency performance by adopting state-of-the-art (SOTA) techniques, such as freezing layers, model distillations, and more. Before we dive into the details, we'll first illustrate the assumptions for the materials of this chapter, as follows:

• We assume there are no...