A scalable NoC topology targeting network performance Avik Bose, Prasun Ghosal Indian Institute of Engineering Science and Technology, Shibpur Howrah, WB, India {avik,p_ghosal}@it.iiests.ac.in ABSTRACT Scalability, path diversity, and bisection-width are the primary de- sign concerns that signifcantly afect the NoC performance. The present work proposes an NoC topology taking into account the above design considerations. Rigorous experimentation is done with prevalent counterparts like mesh, torus, fattened-butterfy, butterfy-fat-tree, etc. Low network diameter and sufcient path diversity come up with a minimum of 38% latency improvement among all the cases. A minimum of 5% throughput yield is also found because of relatively high bisection width. Optimized inter- connect length makes the design fexible to scale with the increasing network size. Relatively lower router radix brings down the average power consumption of the routers signifcantly than the existing topologies. A minimum of 27% gain on average router energy is found. CCS CONCEPTS · Computer systems organization → System on a chip; Mul- ticore architectures; Interconnection architectures; · Networks → Network on chip. KEYWORDS NoC, core layer, bypass layer, path diversity, bridge router. ACM Reference Format: Avik Bose, Prasun Ghosal. 2021. A scalable NoC topology targeting network performance. In 14th edition of International Workshop on Network on Chip Architectures (NoCArc’21), October 18–22, 2021, Virtual Event, Greece. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3477231.3490428 1 INTRODUCTION AND MOTIVATION 1.1 Introduction Due to rapid technology scaling, a large portion of a computing sys- tem can be contained inside a single chip today. The Shift from com- putation to a more communication-centric paradigm has brought Network-on-Chip into the picture [14]. Signifcant research is be- ing carried out to cope with the growing execution demand, which ofers a more efcient NoC infrastructure than before. Handling the trafc load with increasing injection rate in a substantially large network has become a prevalent design challenge that infuences Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specifc permission and/or a fee. Request permissions from permissions@acm.org. NoCArc’21, October 18–22, 2021, Virtual Event, Greece © 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-8711-8/21/10. . . $15.00 https://doi.org/10.1145/3477231.3490428 NoC performance signifcantly [3]. Scalability of the network, along with the existence of signifcant path diversity and bisection width, are the two necessary properties of the underlying network that se- verely impacts the NoC performance [14][3]. Among the prevalent topologies, some designs sufer from large-diameter, whereas oth- ers face long interconnects delay or high router radix (out-degree) problems. The above issues notably impact communication latency, throughput, and power consumption. Therefore, designing a topol- ogy is a critical issue for an NoC architect to achieve sustainable performance. 1.2 Related Research Despite having signifcant path diversity, the diameter of a standard mesh increases rapidly with growing the network size [15][3]. As a result of the above mesh sufers from high communication latency and saturates early at high injection load. Wang et al. demonstrated in their literature [16] that incorporating diagonal links to the routers of a mesh can improve performance as this arrests the rapid growth of diameter to a notable extent. A considerable gain in la- tency and power consumption is observed in [16]. On the other hand, Prasad et al propose the ZMesh [15] network, optimizing the diagonal links of DMesh [16] based on De Bruijn logic which decreases the wire length and achieves better energy efciency com- pared to the counterpart. The existence of ring across both  and  dimensions makes torus perform relatively well than mesh [3]. However, with the growing network size, it sufers from long inter- connect length issues. Butterfy-Fat-Tree (BFT) [5][3] and Flattened Butterfy (Flatfy) [11] converge to saturation later than the above networks but, they face problems with overall power consumption. Despite the high integration density and low network diameter, BFT sufers from long interconnect length, making it vulnerable to implementation after a specifc network size. Kim et al. showcased how Flatfy topology exhibits efcient load balancing and improves latency and throughput compared to mesh. However, the overall power budget of Flatfy considerably increases as it possesses high radix routers. The proposed work attempts to fnd a viable solution taking all the above design considerations into account. 1.3 Novel Contribution The proposed work presents a new topology focusing on the criti- cal design constraints like network diameter, interconnect length, and router radix. Along with the core layer(s), the design uses an additional bypass layer to alleviate the trafc load at a high in- jection rate. Signifcant path diversity and high bisection width give a sustainable performance in substantially large network size. The network is fexible to scale in both 2D and 3D form without afecting the interconnect length and routers’ radix. Floorplanning NoCArc 2021 — October 22, 2021, Global Online Event, Broadcast from Athens, Greece 4