Data volume sculptor for deep learning acceleration

Abstract

The disclosure relates to a data volume sculptor for deep learning acceleration. Embodiments of a device include on-board memory, an applications processor, a digital signal processor (DSP) cluster, aconfigurable accelerator framework (CAF), and at least one communication bus architecture. The communication bus communicatively couples the applications processor, the DSP cluster, and the CAF to the on-board memory. The CAF includes a reconfigurable stream switch and a data volume sculpting unit, which has an input and an output coupled to the reconfigurable stream switch. The data volume sculpting unit has a counter, a comparator, and a controller. The data volume sculpting unit is arranged to receive a stream of feature map data that forms a three-dimensional (3D) feature map. The 3D feature map is formed as a plurality of two-dimensional (2D) data planes. The data volume sculpting unit is also configured to identify a 3D volume within the 3D feature map that is dimensionally smallerthan the 3D feature map and isolate data from the 3D feature map that is within the 3D volume for processing in a deep learning algorithm.

Description

technical field [0001] The present disclosure generally relates to improving flexibility, data locality, and faster execution of deep machine learning systems, such as in convolutional neural networks (CNNs). More specifically, but not exclusively, this disclosure relates to a data volume sculptor for a deep learning acceleration engine. Background technique [0002] Known computer vision, speech recognition and signal processing applications benefit from the use of learning machines. Learning machines discussed in this disclosure may fall under the technical headings of machine learning, artificial intelligence, neural networks, probabilistic inference engines, accelerators, and the like. Such machines are arranged to rapidly perform hundreds, thousands, and millions of concurrent operations. Conventional learning machines can deliver hundreds of teraflops (i.e., a trillion (10 12 ) computing power of floating-point operations). [0003] In some cases, learning machines...

AI Technical Summary

Problems solved by technology

[0067] The eight-layer depth of the AlexNet architecture appears to be important, as certain tests revealed that removing any convolutional layer resulted in unacceptably impaired performance

The size of the network is limited by the amount of memory available on the implemented GPU and by the amount of training time considered tolerable

Images