Hardware accelerator, system and method for accelerating graph neural network attribute access

Abstract

The present disclosure provides a hardware accelerator, a system and a method for accelerating graph neural network attribute access. The hardware accelerator includes a graph neural network attribute processor and a first memory, the graph neural network attribute processor configured in the following way: receiving a first graph node identifier; determining a target memory address within the first memory based on the first graph node identifier; determining whether attribute data corresponding to the received first graph node identifier is cached in a target memory address of a first memory based on the received first graph node identifier; and in response to determining that the attribute data is not cached in the first memory, acquiring the attribute data from the second memory, and writing the acquired attribute data into the target memory address of the first memory. The present disclosure improves memory access efficiency for obtaining graph node attribute data, and thus improves overall performance of graph neural network computations.

Description

technical field [0001] The present disclosure generally relates to accelerating graph neural network (graph neural network, GNN) computation, and in particular, to a hardware accelerator, computer system and method for accelerating graph neural network computation using a multi-level attribute cache mechanism. Background technique [0002] Although traditional deep learning models are good at pattern recognition and data mining by capturing the hidden patterns of European-style data (such as images, texts, videos, etc.), graph neural networks (GNNs) have proven to extend the capabilities of machine learning to Non-European domain of graph data representations with complex relationships and interdependencies between objects. Studies have shown that graph neural networks can surpass the current state-of-the-art in applications ranging from molecular inference to community detection. One of the challenges that has so far hindered the widespread adoption of graph neural network...

AI Technical Summary

Problems solved by technology

One of the challenges that has so far hindered the widespread adoption of graph neural networks in industrial applications is the difficulty of scaling them to large graphs

Images