The increase in the number of cores that can be integrated on a single chip has forced the designer to use computer network concepts for design of System on Chip (SoC). This idea led to development of Network on Chip (NoC) to deal with more cores on a single chip. NoC has three main parts, namely routers, link and network interface through which cores are connected to NoC. Router is one of the most important parts because cores communicate with other cores through routers. One of the important tasks for a NoC designer is to design router with low latency.Router design depends on the routing protocol and routing algorithm used. Two kinds of routing algorithms are source routing and distributed routing. In source routing, complete route information is available in Head flit while in distributed routing, routing decisions are taken inside every router on the path. Source routing has speed advantage over distributed routing because the packet itself contains the routing information. But source routing leads to overhead to store complete path information in the header of each packet. To overcome this flaw, junction based source routing has been introduced. If destination is far away from the source then first packet will go to a junction and get the new path information from the junction to the destination. Thus we need to store the path information only for a few hops in the packet header. This idea has been taken from the daily experience of train journey. In this thesis we have developed design of a router for junction based source routing. Main component of simple router includes buffering, header modification and making route decision. Router includes a table called Path Table which stores information about paths from junction to various destinations. JB router also includes, picking up the new path information from Path Table and modify the header by adding new path information.We have developed VHDL designs of two versions of the routers for Junction Based Routing. The delay performance of routers have been analysed through simulation. A simple prototype of the router has also been implemented in Altera FPGA to find out the resource requirements of the new router designs.

L'utilisation de la troisième dimension peut entraîner une réduction significative de la puissance et de la latence moyenne du trafic dans les réseaux sur puce (Network-on-Chip). La technologie des vias à travers le substrat (ou Through-Silicon Via) est la technologie la plus prometteuse pour l'intégration 3D, car elle offre des liens verticaux courts qui remédient au problème des longs fils dans les NoCs-2D. Les TSVs sont cependant énormes et les processus de fabrication sont immatures, ce qui réduit le rendement des systèmes sur puce à base de NoC-3D. Par conséquent, l'idée de réseaux sur puce 3D partiellement connectés verticalement a été introduite pour bénéficier de la technologie 3D tout en conservant un haut rendement. En outre, de tels réseaux sont flexibles, car le nombre, l'emplacement et l'affectation des liens verticaux dans chaque couche peuvent être décidés en fonction des exigences de l'application. Cependant, ce type de réseaux pose un certain nombre de défis : Le routage est le problème majeur, car l'élimination de certains liens verticaux fait que l'on ne peut utiliser les algorithmes classiques qui suivent l'ordre des dimensions. Pour répondre à cette question nous expliquons et évaluons un algorithme de routage déterministe appelé “Elevator First”, qui garanti d'une part que si un chemin existe, alors on le trouve, et que d'autre part il n'y aura pas d'interblocages. Fondamentalement, la performance du NoC est affecté par a) la micro architecture des routeurs et b) l'architecture d'interconnexion. L'architecture du routeur a un effet significatif sur la performance du NoC, à cause de la latence qu'il induit. Nous présentons la conception et la mise en œuvre de la micro-architecture d'un routeur à faible latence implantant​​l'algorithme de routage Elevator First, qui consomme une quantité raisonnable de surface et de puissance. Du point de vue de l'architecture, le nombre et le placement des liens verticaux ont un rôle important dans la performance des réseaux 3D partiellement connectés verticalement, car ils affectent le nombre moyen de sauts et le taux d'utilisation des FIFOs dans le réseau. En outre, l'affectation des liens verticaux vers les routeurs qui n'ont pas de ports vers le haut ou/et le bas est une question importante qui influe fortement sur les performances. Par conséquent, l'exploration architecturale des réseaux sur puce 3D partiellement connectés verticalement est importante. Nous définissons, étudions et évaluons des paramètres qui décrivent le comportement du réseau, de manière à déterminer le placement et l'affectation des liens verticaux dans les couches de manière simple et efficace. Nous proposons une méthode d'estimation quadratique visantà anticiper le seuil de saturation basée sur ces paramètres. / Utilization of the third dimension can lead to a significant reduction in power and average hop-count in Networks- on-Chip (NoC). TSV technology, as the most promising technology in 3D integration, offers short and fast vertical links which copes with the long wire problem in 2D NoCs. Nonetheless, TSVs are huge and their manufacturing process is still immature, which reduces the yield of 3D NoC based SoC. Therefore, Vertically-Partially-Connected 3D-NoC has been introduced to benefit from both 3D technology and high yield. Moreover, Vertically-Partially-Connected 3D-NoC is flexible, due to the fact that the number, placement, and assignment of the vertical links in each layer can be decided based on the limitations and requirements of the design. However, there are challenges to present a feasible and high-performance Vertically-Partially-Connected Mesh-based 3D-NoC due to the removed vertical links between the layers. This thesis addresses the challenges of Vertically-Partially-Connected Mesh-based 3D-NoC: Routing is the major problem of the Vertically-Partially-Connected 3D-NoC. Since some vertical links are removed, some of the routers do not have up or/and down ports. Therefore, there should be a path to send a packet to upper or lower layer which obviously has to be determined by a routing algorithm. The suggested paths should not cause deadlock through the network. To cope with this problem we explain and evaluate a deadlock- and livelock-free routing algorithm called Elevator First. Fundamentally, the NoC performance is affected by both 1) micro-architecture of routers and 2) architecture of interconnection. The router architecture has a significant effect on the performance of NoC, as it is a part of transportation delay. Therefore, the simplicity and efficiency of the design of NoC router micro architecture are the critical issues, especially in Vertically-Partially-Connected 3D-NoC which has already suffered from high average latency due to some removed vertical links. Therefore, we present the design and implementation the micro-architecture of a router which not only exactly and quickly transfers the packets based on the Elevator First routing algorithm, but it also consumes a reasonable amount of area and power. From the architecture point of view, the number and placement of vertical links have a key role in the performance of the Vertically-Partially-Connected Mesh-based 3D-NoC, since they affect the average hop-count and link and buffer utilization in the network. Furthermore, the assignment of the vertical links to the routers which do not have up or/and down port(s) is an important issue which influences the performance of the 3D routers. Therefore, the architectural exploration of Vertically-Partially-Connected Mesh-based 3D-NoC is both important and non-trivial. We define, study, and evaluate the parameters which describe the behavior of the network. The parameters can be helpful to place and assign the vertical links in the layers effectively. Finally, we propose a quadratic-based estimation method to anticipate the saturation threshold of the network's average latency.