Simulation-Driven Evaluation of Chiplet-Based Architectures Using VisualSim

Wajid Ali, Ayaz Akram, Deepak Shankar · 2025

Read Simulation-Driven Evaluation of Chiplet-Based Architectures Using VisualSim by Wajid Ali, Ayaz Akram, Deepak Shankar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01244.txt.

Introduction

Wajid Ali, Ayaz Akram, and Deepak Shankar simulate chiplet based processors, designs that assemble several small silicon dies into one package instead of building a single large chip. Chiplets promise flexibility and lower cost as monolithic chips get harder to manufacture, but their performance depends on how the pieces communicate.

Using the VisualSim tool, they built a detailed model of a multi die system with multicore ARM processor clusters joined by an ARM network on chip. The simulation measures inter chiplet communication latency, memory access efficiency, workload distribution, and the power performance trade off under varied workloads. The authors use the results to identify the factors that most influence chiplet system performance and to guide later designs.

Catalog record

CREATORWajid Ali, Ayaz Akram, Deepak Shankar
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE13 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-01244

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Simulation-Driven Evaluation of Chiplet-Based Architectures Using VisualSim Wajid Ali, Ayaz Akram, Deepak Shankar Abstract This paper focuses on the simulation of multi-die System-on-Chip (SoC) architectures using VisualSim, emphasizing chiplet-based system modeling and performance analysis. Chiplet technology presents a promising alternative to traditional monolithic chips, which face increasing challenges in manufacturing costs, power efficiency, and performance scaling. By integrating multiple small modular silicon units into a single package, chiplet-based architectures offer greater flexibility and scalability at a lower overall cost. In this study, we developed a detailed simulation model of a chiplet-based system, incorporating multicore ARM processor clusters interconnected through a ARM CMN600 network-on-chip (NoC) for efficient communication [ 4 , 7 ] . The simulation framework in VisualSim enables the evaluation of critical system metrics, including inter-chiplet communication latency, memory access efficiency, workload distribution, and the power-performance tradeoff under various workloads. Through simulation-driven insights, this research highlights key factors influencing chiplet system performance and provides a foundation for optimizing future chiplet-based semiconductor designs. I Introduction The semiconductor industry is facing significant challenges as traditional single-chip designs struggle to meet increasing demands for power, efficiency, and scalability. Moore’s Law, which predicted regular improvements in chip performance, is slowing, and Dennard Scaling, which helped improve performance by shrinking transistors, is no longer effective. Additionally, manufacturing costs are rising, and adding more features to a single chip has become increasingly complex. These trends signal the end of Moore’s Law and mark the beginning of new approaches in computing, such as the introduction of advanced architectures and technologies like chiplets [ 9 ] . Chiplets are small pieces of silicon that can be combined in one package to make a bigger and more flexible System-on-Chip (SoC). Instead of using one big chip, chiplet-based designs use smaller parts that are made for specific tasks. These parts communicate with each other using advanced connection methods like the Universal Chiplet Interconnect Express (UCIe) [ 8 ] . This approach offers several benefits, such as reduced manufacturing costs, improved yields, and the ability to design chips tailored for specific applications [ 5 , 6 , 3 ] . Chiplets also provide scalability, as additional units can be integrated to handle more tasks without redesigning the entire system. However, designing and optimizing chiplet-based systems presents challenges. Engineers must consider memory operations, inter-chiplet communication, power usage, and thermal management to achieve efficient performance [ 5 , 6 ] . To handle these challenges, tools like VisualSim Architect are very important. VisualSim Architect is a platform that helps engineers model and test complex chip designs in a virtual environment. It allows engineers to create detailed models of chip components, simulate how the system will behave under different conditions, and study important results like speed, power consumption, data transfer efficiency, and heat generation. By using VisualSim, engineers can identify potential problems, optimize performance, and make better decisions during the design phase. VisualSim Architect has a user-friendly graphical interface, making it accessible for both beginners and experts. It also includes a large library of ready-made components, such as processors, memory units, and interconnects, which can be easily combined to build complex Systems-on-Chip (SoC) architectures. For chiplet-based designs, VisualSim allows engineers to study how the chiplets communicate, analyze memory access patterns, and explore different interconnect settings to ensure the system performs efficiently [ 6 , 3 , 1 ] . One of the most important features of VisualSim is its ability to simulate real-world and dynamic workloads. This helps engineers see how their designs will perform under actual usage and different conditions. For example, they can test what happens when more cores are added to a chiplet, analyze memory access patterns, or check how changes in interconnect settings like bandwidth and latency affect performance. These simulations give engineers valuable insights, helping them identify bottlenecks, optimize designs, and make better decisions before manufacturing. This reduces the need for expensive prototypes, saves time, lowers development costs, and ensures efficient, scalable, and reliable chip designs for modern applications [ 6 , 1 ] . In this study, we use VisualSim Architect to design, model, and test a chiplet-based multi-die System-on-Chip (SoC). Each chiplet contains processor clusters that communicate within the chiplet using a CMN600 network-on-chip (NoC) [ 2 ] . The chiplets are connected to each other using UCIe links, which ensure smooth communication between them. Our study focuses on three main goals: (1) studying how different memory setups, like shared DRAM in a single-chip design and separate DRAM modules in a chiplet-based design, affect performance and scalability, (2) examining how adding more cores in chiplets impacts processing power and the balance between performance and energy use, and (3) testing how interconnect technologies, such as UCIe and CMN600 [ 8 , 7 ] , help chiplets communicate and transfer data efficiently in scalable systems. II Methodology This study evaluates a multi-die System-on-Chip (SoC) architecture using VisualSim Architect, a simulation platform designed for detailed modeling and performance analysis. The focus is on understanding how various architectural configurations, including monolithic and chiplet-based designs, perform under real-world workloads. The experiments explore the impact of factors such as memory hierarchy, core distribution, and interconnect technologies on key performance metrics like latency, power consumption, and scalability. By leveraging VisualSim Architect, the study models complex components like CPUs, caches, memory controllers, and interconnects to simulate task execution and data flow within the system. This approach enables detailed analysis of how design changes affect system efficiency, helping optimize both monolithic and chiplet-based SoC designs. The methodology consists of two key phases: experimental setup and simulation framework to assess system performance under varying configurations. II-A Experimental Setup The system is based on the ARM A720AE model, configured with multicore clusters, DRAM modules, and interconnections using CMN600 NoC and UCIe links. In Experiment 1, a single monolithic chip with two ARM A720AE clusters is used. Each cluster contains 1 core with private L1 Data and Instruction Caches, a private L2 Cache, and a shared L3 Cache. Both clusters are connected via a CMN600 Coherent Mesh Network (NoC), and a two DRAM module is used for memory requests from both clusters. This setup allows for scalability to add more cores. CPU CPU CMN DRAM DRAM Fig. 1: Block diagram of Experiment 1: Monolithic SoC In Experiment 2, the system is expanded to a chiplet-based design with two ARM A720AE clusters placed on separate dies. The clusters are connected using UCIe interconnects, and each chiplet retains its private cache hierarchy and a dedicated DRAM module. All other parameters remain the same as in Experiment 1. CPU CPU CMN CMN DRAM DRAM UCIe CHI Fig. 2: Block diagram of Experiment 2: Chiplet-Based SoC. In Experiment 3, the system retains the chiplet-based design but increases the number of cluster from 2 to 4, significantly enhancing computational capacity. Each chiplet is equipped with two DRAM modules connected via CMN600, providing additional memory bandwidth. DRAM CMN CPU CMN DRAM CPU DRAM CMN CPU CMN DRAM CPU UCIe CHI UCIe CHI Fig. 3: Block diagram of Experiment 3: Multi-Chiplet SoC. II-B Simulation Framework VisualSim Architect is used for modeling and simulation. This platform enables the creation of detailed models for system components, including task graphs, CPUs, caches, memory, and interconnects. The simulations analyze

Read the complete text