Chiplets on Wheels : Review Paper on holistic chiplet solutions for autonomous vehicles
By Swathi Narashiman, Venkat A, Divyaratna Joshi, Deepak Sridhar, Harish Rajesh, Sanjay Sattva, Aniruddha S, Jayanth B, Varun Manjunath, and Ragavendiran N (Indian Institute of Technology, Madras)
Abstract
On the advent of the slow death of Moore’s law, the silicon industry is moving towards a new era of chiplets.The automotive industry is experiencing a profound transformation towards software-defined vehicles, fueled by the surging demand for automotive compute chips, expected to reach 20−22 billion by 2030. High-performance compute (HPC) chips become instrumental in meeting the soaring demand for computational power. Various strategies, including centralized electrical and electronic architecture and the innovative Chiplet Systems, are under exploration. The latter, breaking down System-on-Chips (SoCs) into functional units, offers unparalleled customization and integration possibilities. The research accentuates the crucial open Chiplet ecosystem, fostering collaboration and enhancing supply chain resilience. In this paper, We address the unique challenges that arise when we attempt to leverage chiplet-based architecture to design a holistic silicon solution for the automotive industry. We propose a throughput-oriented micro-architecture for ADAS & infotainment system alongside a novel methodology to evaluate chiplet architectures. Further, We develop in-house simulation tools leveraging the gem5 framework to simulate for latency and throughput.Finally, We perform an extensive design of thermally-aware chiplet placement and develop a micro-fluids based cooling design. 1
Keywords: Chiplet Consolidation ⋅ Micro-architecture ⋅ Autonomous compute ⋅ Packaging
1 Introduction
The automotive industry is undergoing a significant transformation as it shifts towards software-defined cars, driven by the growing demand for automotive compute chips. The automotive compute market is expected to grow to $20 billion to $22 billion in 2030 [1]. This shift is leading to a change in focus from hardware to software within the industry. To meet the increasing demand for compute power, high-performance compute (HPC) chips are crucial, as automakers seek a resilient supply chain for customizable System-on-Chips (SoCs), while chip suppliers aim to recoup their upfront research and development (R&D) investments.
Multiple approaches are being considered to transition towards a single compute stack, including a gradual migration to a centralized electrical and electronic (E/E) architecture, employing domain controllers, customized SoCs, and Chiplet Systems. Chiplet Systems involve breaking down SoCs into functional Chiplets, offering greater customization and integration possibilities, ultimately supporting future centralized compute stacks.
An open ecosystem of Chiplet Systems can create a new value chain dynamic within the automotive compute market, which is projected to grow significantly. This open ecosystem enables various stakeholders, including automakers, Tier-1 suppliers, and smaller chip suppliers, to collaborate in chip development, enhancing supply chain resilience and reducing the risk of lock-in effects. Overall, this shift in the automotive compute industry holds the potential for various players to benefit from these new dynamics and opportunities.
In this context, We aim to perform a holistic analysis of adoption of Chiplet technology in the automotive industry. The paper is organized as follows: Section 2 gives an overview of the automotive E&E landscape, evolving compute demands and the challenges ahead. Section 3 first describes the qualitative benefits of chiplets over a SoC and then proceeds to quantitatively evaluate the differences between SoCs and Chiplets in terms of cost, power and Performance. Next We propose 2 areas of application of Chiplets in the automotive compute stack, giving a overview of special purpose computing. Then we give a detailed analysis of leading communication technologies and the interconnect IPs that are used in the industry. Finally We propose a thermal and packaging solution for the chiplet architecture we propose.
We will also explore the cutting-edge technologies in the field of chiplet design, specifically focusing on novel interconnect technologies and communication protocols. Our investigation will compare these advancements to traditional monolithic System-on-Chip (SoC) architectures. To provide a comprehensive analysis, we will delve deeper into various interconnect protocols, examining their latency and communication efficiency. This examination will be supported by simulations as well as insights derived from a synthesis of research papers and technical articles.
In final sections of the paper, We present a thermal design methodology, using simulation tools that optimize the complex dependencies involved in packaging a chip. We design a cooling solution based on microfluidics, giving special considertions to the unique constraints posed by the automotive environment.
Related Chiplet
- Direct Chiplet Interface
- HBM3e Advanced-packaging chiplet for all workloads
- UCIe AP based 8-bit 170-Gsps Chiplet Transceiver
- UCIe based 8-bit 48-Gsps Transceiver
- UCIe based 12-bit 12-Gsps Transceiver
Related Technical Papers
- Stop-For-Top IP model to replace One-Stop-Shop by 2025... and support the creation of successful Chiplet business
- Chiplet Strategy is Key to Addressing Compute Density Challenges
- RapidChiplet: A Toolchain for Rapid Design Space Exploration of Chiplet Architectures
- Chiplet Cloud: Building AI Supercomputers for Serving Large Generative Language Models
Latest Technical Papers
- Performance Implications of Multi-Chiplet Neural Processing Units on Autonomous Driving Perception
- ChipAI: A scalable chiplet-based accelerator for efficient DNN inference using silicon photonics
- Advanced Packaging and Chiplets Can Be for Everyone
- Interfacing silicon photonics for high-density co-packaged optics
- System-Technology Co-Optimization for Dense Edge Architectures using 3D Integration and Non-Volatile Memory