Hardware Acceleration of Graph Neural Networks

Auten, Adam; Tomei, Matthew; Kumar, Rakesh

doi:10.1109/dac18072.2020.9218751

Cited by 90 publications

(55 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Even though advances in sparse/irregular tensor processing [34] and graph processing [63,154] may prove useful in accelerating GNNs, addressing their unique computing challenges requires more specialized proposals. Some attempts have been done from a software perspective, i.e., adapting the GNN operations to better match the capabilities of CPUs or GPUs [106,144,155]; and from a hardware perspective, i.e., designing custom processors tailored to the demands of GNNs [7,53,103,164]. However, recent surveys and reviews [11,16,19,66,91,160,181,185] lack a comprehensive analysis of such advances.…”

Section: Deep Learning On Graphsmentioning

confidence: 99%

Computing Graph Neural Networks: A Survey from Algorithms to Accelerators

et al. 2021

View full text Add to dashboard Cite

Graph Neural Networks (GNNs) have exploded onto the machine learning scene in recent years owing to their capability to model and learn from graph-structured data. Such an ability has strong implications in a wide variety of fields whose data are inherently relational, for which conventional neural networks do not perform well. Indeed, as recent reviews can attest, research in the area of GNNs has grown rapidly and has lead to the development of a variety of GNN algorithm variants as well as to the exploration of ground-breaking applications in chemistry, neurology, electronics, or communication networks, among others. At the current stage research, however, the efficient processing of GNNs is still an open challenge for several reasons. Besides of their novelty, GNNs are hard to compute due to their dependence on the input graph, their combination of dense and very sparse operations, or the need to scale to huge graphs in some applications. In this context, this article aims to make two main contributions. On the one hand, a review of the field of GNNs is presented from the perspective of computing. This includes a brief tutorial on the GNN fundamentals, an overview of the evolution of the field in the last decade, and a summary of operations carried out in the multiple phases of different GNN algorithm variants. On the other hand, an in-depth analysis of current software and hardware acceleration schemes is provided, from which a hardware-software, graph-aware, and communication-centric vision for GNN accelerators is distilled.

show abstract

Section: Deep Learning On Graphsmentioning

confidence: 99%

Computing Graph Neural Networks: A Survey from Algorithms to Accelerators

et al. 2021

View full text Add to dashboard Cite

show abstract

“…Ideally, developing a method to parallelize feature extraction across subareas (as they are independent of one another) may offer a significant improvement to inference time. Future works for accelerating graph processing on the hardware level will also offer a significant improvement of inference time [41].…”

Section: Discussionmentioning

confidence: 99%

Rotation Invariance in Graph Convolutional Networks

Mac¹,

Nguyen²

2021

Annals of Computer Science and Information Systems

View full text Add to dashboard Cite

Convolution filters in deep convolutional networks display rotation variant behavior. While learned invariant behavior can be partially achieved, this paper shows that current methods of utilizing rotation variant features can be improved by proposing a grid-based graph convolutional network. By performing spectral graph convolutions on features extracted from subareas of images, we are able to take advantage of the geometric nature of relational machine learning in graph neural networks to be able to overcome rotation variant features to perform object localization. We demonstrate that Grid-GCN heavily outperforms existing models on rotated images, and through a set of ablation studies, we show how the performance of Grid-GCN implies that there exist more performant methods to utilize fundamentally rotation variant features and we conclude that the inherit nature of spectral graph convolutions is able to learn invariant behavior.

show abstract

“…Outside of HEP, hardware and firmware acceleration of GNN inference, and graph processing in general, has been an active area of study in recent years, motivated by the intrinsic inefficiencies of CPUs and GPUs when dealing with graph data (Besta et al, 2019;Gui et al, 2019). Nurvitadhi et al, 2014;Ozdal et al, 2016;Auten et al, 2020;Geng et al, 2020;Kiningham et al, 2020;Yan et al, 2020;Zeng and Prasanna, 2020 describe examples of GNN acceleration architectures. Auten et al, 2020;Geng et al, 2020;Yan et al, 2020;Zeng and Prasanna, 2020.…”

Section: Related Workmentioning

confidence: 99%

“…Nurvitadhi et al, 2014;Ozdal et al, 2016;Auten et al, 2020;Geng et al, 2020;Kiningham et al, 2020;Yan et al, 2020;Zeng and Prasanna, 2020 describe examples of GNN acceleration architectures. Auten et al, 2020;Geng et al, 2020;Yan et al, 2020;Zeng and Prasanna, 2020. are specific to the graph convolutional network (GCN) (Kipf and Welling, 2017), while the graph inference processor (GRIP) architecture in Kiningham et al, (2020) is efficient across a wide range of GNN models.…”

Section: Related Workmentioning

confidence: 99%

Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics

et al. 2021

View full text Add to dashboard Cite

Graph neural networks have been shown to achieve excellent performance for several crucial tasks in particle physics, such as charged particle tracking, jet tagging, and clustering. An important domain for the application of these networks is the FGPA-based first layer of real-time data filtering at the CERN Large Hadron Collider, which has strict latency and resource constraints. We discuss how to design distance-weighted graph networks that can be executed with a latency of less than one μs on an FPGA. To do so, we consider a representative task associated to particle reconstruction and identification in a next-generation calorimeter operating at a particle collider. We use a graph network architecture developed for such purposes, and apply additional simplifications to match the computing constraints of Level-1 trigger systems, including weight quantization. Using the hls4ml library, we convert the compressed models into firmware to be implemented on an FPGA. Performance of the synthesized models is presented both in terms of inference accuracy and resource usage.

show abstract

Hardware Acceleration of Graph Neural Networks

Cited by 90 publications

References 11 publications

Computing Graph Neural Networks: A Survey from Algorithms to Accelerators

Computing Graph Neural Networks: A Survey from Algorithms to Accelerators

Rotation Invariance in Graph Convolutional Networks

Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics

Contact Info

Product

Resources

About