Tensor Accelerator Architecture

In Part VI, we move from general-purpose architectures to specialized hardware structures known as tensor accelerators. These structures give up flexibility to spend their silicon (and energy) budget catering to specific patterns within a class of workloads and offer a massive performance boost on those patterns. Since these architectures must be tuned for specific types of workloads, they are fast-evolving and far more heterogeneous. Designs and feature sets differ between vendors and generations, and publicly available information can be sparse. Hence, the treatment in this chapter is correspondingly broader and shallower than other similar chapters, such as 11 (CPU) and 14 (GPU). The essential goal is to build a mental map that makes the hands-on work in Chapter 17 intuitive and to develop a vocabulary for self-guided research.