Fusion memory, storage system and deep learning calculation method

Abstract

The embodiment of the invention provides a fusion memory, a storage system and a deep learning calculation method. The fusion memory comprises an acceleration storage controller and a storage chip; the acceleration storage controller comprises a processor, a data path module and an acceleration processing module; the processor is used for scheduling the data path module according to the received storage command and storing externally input data to the storage chip corresponding to the storage command, and the data comprises to-be-calculated data; the processor is used for scheduling a data path module according to the obtained calculation control instruction and a loading mode corresponding to the calculation control instruction, and loading the to-be-calculated data to an acceleration processing module; the data path module is used for providing a transmission path for stored data and providing a transmission path for loading to-be-calculated data; the acceleration processing module is used for performing deep learning calculation on the to-be-calculated data through hardware acceleration; and the storage chip is used for storing data. Through the scheme, the calculation efficiency of deep learning calculation can be improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a fusion memory, a storage system and a deep learning calculation method. Background technique [0002] At present, the hierarchical coupling architecture of computing nodes and storage nodes is widely used to perform deep learning calculations on data, such as figure 1 shown. In the computing node, AU (Acceleration Unit, accelerated processing module) for deep learning is deployed. AU is a chip that realizes deep learning calculation through hardware acceleration, such as TPU (Tensor Processing Unit, tensor processor), GPU (Graphics Processing Unit, image processor), etc., AU as a coprocessor is responsible for scheduling by the CPU (Central Processing Unit, central processing unit) in the computing node; the data to be calculated required for deep learning computing is stored in the storage node. [0003] In the above layered coupling architecture of computing...

AI Technical Summary

Problems solved by technology

[0004] Due to the movement of the data to be calculated between the storage node and the computing node, it is easily affected by the external environment, and there are inevitably problems of delay and packet loss during the transmission of the data to be calculated, resulting in low computational efficiency of deep learning calculations

Images