Titlebook: Efficient Processing of Deep Neural Networks; Vivienne Sze,Yu-Hsin Chen,Joel S. Emer Book 2020 Springer Nature Switzerland AG 2020

BRINK

Designing DNN Acceleratorsmultiplications, in order to achieve higher performance (i.e., higher throughput and/or lower latency) on off-the-shelf general-purpose processors such as CPUs and GPUs. In this chapter, we will focus on optimizing the processing of DNNs directly by designing specialized hardware.

Liberate

Customary

intuition

Exploiting Sparsityeferring to the fact that there are many repeated values in the data. Much of the time the repeated value is zero, which is what we will assume unless explicitly noted. Thus, we will talk about the sparsity or density of the data as the percentage of zeros or non-zeros, respectively in the data. The

delegate

Pepsin

Advanced Technologiess well as the transfer of the data. The associated physical factors also limit the bandwidth available to deliver data between memory and compute, and thus limits the throughput of the overall system. This is commonly referred to by computer architects as the “memory wall.”

Ingredient

Conclusionations including computer vision, speech recognition, and robotics and are often delivering better than human accuracy. However, while DNNs can deliver this outstanding accuracy, it comes at the cost of high computational complexity. With the stagnation of improvements in general-purpose computation

enflame

Bumptious

富饒

https://doi.org/10.1007/978-3-662-39613-1stical learning on a large amount of data to obtain an effective representation of an input space. This is different from earlier approaches that use hand-crafted features or rules designed by experts.