Tampere University of Technology

TUTCRIS Research Portal

On design and comparison of on-chip networks

Research output: Book/ReportDoctoral thesisMonograph

Standard

On design and comparison of on-chip networks. / Salminen, Erno.

Tampere : Tampere University of Technology, 2010. (Tampereen teknillinen yliopisto. Julkaisu; Vol. 872).

Research output: Book/ReportDoctoral thesisMonograph

Harvard

Salminen, E 2010, On design and comparison of on-chip networks. Tampereen teknillinen yliopisto. Julkaisu, vol. 872, Tampere University of Technology, Tampere.

APA

Salminen, E. (2010). On design and comparison of on-chip networks. (Tampereen teknillinen yliopisto. Julkaisu; Vol. 872). Tampere: Tampere University of Technology.

Vancouver

Salminen E. On design and comparison of on-chip networks. Tampere: Tampere University of Technology, 2010. (Tampereen teknillinen yliopisto. Julkaisu).

Author

Salminen, Erno. / On design and comparison of on-chip networks. Tampere : Tampere University of Technology, 2010. (Tampereen teknillinen yliopisto. Julkaisu).

Bibtex - Download

@book{423514e66df148c7bd4be6484219d9db,
title = "On design and comparison of on-chip networks",
abstract = "This thesis focuses on the design of on-chip communication networks and methods for benchmarking them. Network-on-Chip (NoC) paradigm seeks to achieve greater design productivity and performance in large integrated circuits. Such systems include heterogeneous set of components that have different requirements for communication. This thesis presents simulation-based evaluation methods for NoCs. In addition, several detailed guidelines are given in order to promote disciplined NoC benchmarking. Discussion starts with thorough surveys of 60 existing NoCs and over 40 evaluation studies. The presented benchmarking methodology relies on abstract workload models based on task graphs. They are executed with a Transaction Generator (TG) that sends and receives data to/from the benchmarked NoC and collects statistics. TG was used in several configurations and was essential part for completing this work. The error in time estimates was mostly below 10{\%} whereas the speedup against cycleaccurate HW/SW co-simulation was over 200x. Heterogeneous IP Block Interconnection (HIBI) was designed to obtain a topologyindependent, scalable, and still high-performance network for integrating intellectual property blocks. Six other NoCs were implemented for reference and benchmarked with HIBI in various configurations and using multiple workloads. Over 30 published implementation results were gathered. Furthermore, several FPGA prototypes were implemented and they confirmed the utility of HIBI in multiprocessor environment. In general, HIBI and 2-D mesh performed better than others in the presented cases considering the trade-off between area and throughput. The main goals of the work were met. The presented methodology along with TG has been adopted by an OCP-IP workgroup that is seeking to standardize NoC benchmarking methods.",
author = "Erno Salminen",
note = "Awarding institution:Tampere University of Technology",
year = "2010",
language = "English",
isbn = "978-952-15-2825-0",
series = "Tampereen teknillinen yliopisto. Julkaisu",
publisher = "Tampere University of Technology",

}

RIS (suitable for import to EndNote) - Download

TY - BOOK

T1 - On design and comparison of on-chip networks

AU - Salminen, Erno

N1 - Awarding institution:Tampere University of Technology

PY - 2010

Y1 - 2010

N2 - This thesis focuses on the design of on-chip communication networks and methods for benchmarking them. Network-on-Chip (NoC) paradigm seeks to achieve greater design productivity and performance in large integrated circuits. Such systems include heterogeneous set of components that have different requirements for communication. This thesis presents simulation-based evaluation methods for NoCs. In addition, several detailed guidelines are given in order to promote disciplined NoC benchmarking. Discussion starts with thorough surveys of 60 existing NoCs and over 40 evaluation studies. The presented benchmarking methodology relies on abstract workload models based on task graphs. They are executed with a Transaction Generator (TG) that sends and receives data to/from the benchmarked NoC and collects statistics. TG was used in several configurations and was essential part for completing this work. The error in time estimates was mostly below 10% whereas the speedup against cycleaccurate HW/SW co-simulation was over 200x. Heterogeneous IP Block Interconnection (HIBI) was designed to obtain a topologyindependent, scalable, and still high-performance network for integrating intellectual property blocks. Six other NoCs were implemented for reference and benchmarked with HIBI in various configurations and using multiple workloads. Over 30 published implementation results were gathered. Furthermore, several FPGA prototypes were implemented and they confirmed the utility of HIBI in multiprocessor environment. In general, HIBI and 2-D mesh performed better than others in the presented cases considering the trade-off between area and throughput. The main goals of the work were met. The presented methodology along with TG has been adopted by an OCP-IP workgroup that is seeking to standardize NoC benchmarking methods.

AB - This thesis focuses on the design of on-chip communication networks and methods for benchmarking them. Network-on-Chip (NoC) paradigm seeks to achieve greater design productivity and performance in large integrated circuits. Such systems include heterogeneous set of components that have different requirements for communication. This thesis presents simulation-based evaluation methods for NoCs. In addition, several detailed guidelines are given in order to promote disciplined NoC benchmarking. Discussion starts with thorough surveys of 60 existing NoCs and over 40 evaluation studies. The presented benchmarking methodology relies on abstract workload models based on task graphs. They are executed with a Transaction Generator (TG) that sends and receives data to/from the benchmarked NoC and collects statistics. TG was used in several configurations and was essential part for completing this work. The error in time estimates was mostly below 10% whereas the speedup against cycleaccurate HW/SW co-simulation was over 200x. Heterogeneous IP Block Interconnection (HIBI) was designed to obtain a topologyindependent, scalable, and still high-performance network for integrating intellectual property blocks. Six other NoCs were implemented for reference and benchmarked with HIBI in various configurations and using multiple workloads. Over 30 published implementation results were gathered. Furthermore, several FPGA prototypes were implemented and they confirmed the utility of HIBI in multiprocessor environment. In general, HIBI and 2-D mesh performed better than others in the presented cases considering the trade-off between area and throughput. The main goals of the work were met. The presented methodology along with TG has been adopted by an OCP-IP workgroup that is seeking to standardize NoC benchmarking methods.

M3 - Doctoral thesis

SN - 978-952-15-2825-0

T3 - Tampereen teknillinen yliopisto. Julkaisu

BT - On design and comparison of on-chip networks

PB - Tampere University of Technology

CY - Tampere

ER -