Nuevas arquitecturas hardware de procesamiento de alto rendimiento para aprendizaje profundo

The design and manufacture of hardware is expensive, both in time and in economic investment, which is why integrated circuits are always manufactured in large volume, to take advantage of economies of scale. For this reason, the majority of processors manufactured are general purpose, thus expanding its range of applications. In recent years, however, more and more processors are being manufactured for specific applications, including those aimed at accelerating work with deep neural networks. This article introduces the need for this type of specialized hardware, describing its purpose, operation and current implementations.