A Scalable and Low-Power FPGA-Aware Network-on-Chip Architecture

Somnath Mazumdar*, Alberto Scionti, Antoni Portero, Jan Martinovič, Olivier Terzo

*Corresponding author af dette arbejde

Publikation: Bidrag til bog/antologi/rapportKonferencebidrag i proceedingsForskningpeer review

Abstrakt

The growing demand for high-performance capabilities in data centers (DCs) leads to adopt heterogeneous solutions. The advantage of specialised hardware is a better support for different types of workloads, and a reduction of the power consumption. Among the others, FPGAs offer the unique capability to provide hardware specialisation and low power consumption. In this context, large arrays of simple and reconfigurable processing elements (PEs), known as coarse-grain reconfigurable arrays (CGRAs), represent a flexible solution for supporting heterogeneous workloads through a specialised instruction set that provides high performance in specific application domains (e.g., image recognition, patterns classification). However, efficient and scalable interconnections are required to sustain throughput and performance of CGRAs. To this end, networks-on-chip (NoCs) have been recognised as a viable solution for better data packet communication. In this paper, we propose an FPGA-aware NoC design targeting CGRAs with 128+ PEs. The proposed design leverages on a two-level topology to scale well with the increasing number of PEs, while the introduction of a software-defined reconfiguration capability offers the opportunity to tailor the set of resources assigned to a specific application. Partitions of physical resources (i.e., virtual domains) are built over the physical topology to meet the required performance, as well as to ease sharing physical chip resources among applications. Experimental evaluation shows the efficiency of our solution regarding used FPGA resources and power consumption.
OriginalsprogEngelsk
TitelComplex, Intelligent, and Software Intensive Systems : Proceedings of the 11th International Conference on Complex, Intelligent, and Software Intensive Systems, CISIS 2017
RedaktørerLeonard Barolli, Olivier Terzo
Antal sider14
UdgivelsesstedCham
ForlagSpringer VS
Publikationsdato2018
Sider407-420
ISBN (Trykt)9783319615653
ISBN (Elektronisk)9783319615660
DOI
StatusUdgivet - 2018
Udgivet eksterntJa
Begivenhed11th International Conference on Complex, Intelligent, and Software Intensive Systems, CISIS 2017 - Torino, Italien
Varighed: 10 jul. 201712 jul. 2017
Konferencens nummer: 11

Konference

Konference11th International Conference on Complex, Intelligent, and Software Intensive Systems, CISIS 2017
Nummer11
Land/OmrådeItalien
ByTorino
Periode10/07/201712/07/2017
NavnAdvances in Intelligent Systems and Computing
Vol/bind611
ISSN2194-5357

Citationsformater