Characterizing the Communication Requirements of GNN Accelerators: A Model-Based Approach

Guirado, Robert; Jain, Akshay; Abadal, Sergi; Alarcón, Eduard

doi:10.1109/iscas51556.2021.9401612

Cited by 7 publications

(3 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Finally, considering the critical environmental sustainability target in 6G, AI/ML can be used to incorporate prediction capabilities when implementing energy-aware network management to be more energy efficient and even achieving further energy savings by means of taking proactive decisions. But at the same time AI/ML would be bringing its own energy consumption increase, especially on the ML training phase, when the significant computing process takes place [93]. Therefore, the optimal balance between reached performance and energy consumption should be found.…”

Section: ) Enablers For Closed-loop Control and Automationmentioning

confidence: 99%

Design Methodology for 6G End-to-End System: Hexa-X-II Perspective

Kerboeuf,

Porambage,

Jain

et al. 2024

IEEE Open J. Commun. Soc.

Self Cite

View full text Add to dashboard Cite

As the fifth generation (5G) mobile communication systems are commercially deployed, they bring new services, enhance user experiences, and offer various opportunities to different industries. Despite its advancements, 5G encounters several challenges. To tackle these issues, global industrial, academic, and standards organizations are actively researching on sixth generation (6G) wireless communication systems. 6G networks are envisioned as a transformative shift in the interactions between the physical, digital, and human realms, paving the way for a pervasive human-centered cyber-physical world. 6G is positioned to be a platform that offers communication and beyond communication services considering both performance and value centric technological development approaches. In this paper, we present the way forward towards the design of 6G end-to-end (E2E) system as a consolidated view of leading industry stakeholders and academia in the European level 6G flagship project Hexa-X-II. We discuss the key challenges with 5G and how 6G is expected to tackle those by introducing new technological innovations and supporting novel use case requirements. We provide a comprehensive methodology for the design of a 6G E2E system including ten principles, a blueprint, and a structured design process. The architecture design principles prioritize environmental sustainability, digital inclusiveness, and trustworthiness, considering their impact on the 6G E2E system. The blueprint is described corresponding to the infrastructure, network centric application, and application layers, as well as the pervasive functionalities and the relevant technological innovations. Following the design principles and the system blueprint, the design process is demonstrated as two-way approaches (i.e., 1) key performance and value indicators based design process. 2) top-down versus bottomup alignment process) through the lens of a collaborative robot use case. Through this use case, a special attention is given to the technological enablers that cover management and orchestration functionalities and their 6G enhancement to go beyond the limitations characterizing the previous network generation. In addition, virtual modelling aspects related to digital twining and simulations for 6G E2E system design are also discussed. INDEX TERMS 6G, end-to-end, platform, system, blueprint, network architecture, design process, use case, key value indicators, key performance indicators, collaborative robots.

show abstract

Section: ) Enablers For Closed-loop Control and Automationmentioning

confidence: 99%

Design Methodology for 6G End-to-End System: Hexa-X-II Perspective

Kerboeuf,

Porambage,

Jain

et al. 2024

IEEE Open J. Commun. Soc.

Self Cite

View full text Add to dashboard Cite

show abstract

“…Several GNN accelerator ASICs have been proposed [19]- [24], [28]- [30], each implementing a specific dataflow which is heavily co-designed with the microarchitecture. The dataflows from these accelerators form a subset of the designspace that we explore.…”

Section: B Gnn Acceleratorsmentioning

confidence: 99%

Understanding the Design-Space of Sparse/Dense Multiphase GNN dataflows on Spatial Accelerators

Garg¹,

Qin²,

Muñoz-Martínez³

et al. 2022

2022 IEEE International Parallel and Distributed Processing Symposium (IPDPS)

Self Cite

View full text Add to dashboard Cite

Graph Neural Networks (GNNs) have garnered a lot of recent interest because of their success in learning representations from graph-structured data across several critical applications in cloud and HPC. Owing to their unique compute and memory characteristics that come from an interplay between dense and sparse phases of computations, the emergence of reconfigurable dataflow (aka spatial) accelerators offers promise for acceleration by mapping optimized dataflows (i.e., computation order and parallelism) for both phases. The goal of this work is to characterize and understand the design-space of dataflow choices for running GNNs on spatial accelerators in order for mappers or design-space exploration tools to optimize the dataflow based on the workload. Specifically, we propose a taxonomy to describe all possible choices for mapping the dense and sparse phases of GNN inference, spatially and temporally over a spatial accelerator, capturing both the intra-phase dataflow and the inter-phase (pipelined) dataflow. Using this taxonomy, we do deep-dives into the cost and benefits of several dataflows and perform case studies on implications of hardware parameters for dataflows and value of flexibility to support pipelined execution.

show abstract

“…Recently, GCNAX [26] proposed a flexible accelerator with Design Space Exploration to choose the ideal dataflow for a workload. Recently, there has also been work on analytical modelling of data movement for GNN accelerators [12]. VII.…”

Section: Related Workmentioning

confidence: 99%

A Taxonomy for Classification and Comparison of Dataflows for GNN Accelerators

Garg¹,

Qin²,

Muñoz-Martínez³

et al. 2021

Self Cite

View full text Add to dashboard Cite

Recently, Graph Neural Networks (GNNs) have received a lot of interest because of their success in learning representations from graph structured data. However, GNNs exhibit different compute and memory characteristics compared to traditional Deep Neural Networks (DNNs). Graph convolutions require feature aggregations from neighboring nodes (known as the aggregation phase), which leads to highly irregular data accesses. GNNs also have a very regular compute phase that can be broken down to matrix multiplications (known as the combination phase). All recently proposed GNN accelerators utilize different dataflows and microarchitecture optimizations for these two phases. Different communication strategies between the two phases have been also used. However, as more custom GNN accelerators are proposed, the harder it is to qualitatively classify them and quantitatively contrast them. In this work, we present a taxonomy to describe several diverse dataflows for running GNN inference on accelerators. This provides a structured way to describe and compare the design-space of GNN accelerators.

show abstract

Characterizing the Communication Requirements of GNN Accelerators: A Model-Based Approach

Cited by 7 publications

References 20 publications

Design Methodology for 6G End-to-End System: Hexa-X-II Perspective

Design Methodology for 6G End-to-End System: Hexa-X-II Perspective

Understanding the Design-Space of Sparse/Dense Multiphase GNN dataflows on Spatial Accelerators

A Taxonomy for Classification and Comparison of Dataflows for GNN Accelerators

Contact Info

Product

Resources

About