CarbonPATH: Carbon-aware pathfinding and architecture optimization for chiplet-based AI systems
Chetan Choppali Sudarshan, Jiajun Hu, Aman Arora, Vidya A. Chhabria · 2026
Read CarbonPATH: Carbon-aware pathfinding and architecture optimization for chiplet-based AI systems by Chetan Choppali Sudarshan, Jiajun Hu, Aman Arora, Vidya A. Chhabria (2026) 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-2603-03878.txt.
Introduction
Chetan Choppali Sudarshan and others present CarbonPATH, a framework for designing chiplet-based AI systems with carbon footprint as a first-class concern. AI growth is driving demand for computing while enlarging its footprint on the environment. Makers are shifting to heterogeneous integration, building systems from chiplets, to escape the yield and cost limits of single-die chips. Searching this space means balancing performance, cost, and carbon across apps, architecture, chiplets, and packaging.
CarbonPATH explores the space with simulated annealing, co-designing workload mapping, design settings, and packaging methods. It accounts for compute and memory sizes, chiplet process nodes, wiring protocols, integration styles from flat 2.5D layouts to stacked 3D ones, running and embodied carbon, and link type. By capturing interactions across levels, it finds system-level solutions that narrower methods miss, supporting earlier, carbon-aware decisions in AI hardware design.
Catalog record
| CREATOR | Chetan Choppali Sudarshan, Jiajun Hu, Aman Arora, Vidya A. Chhabria |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 92788 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03878
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CarbonPATH: Carbon-aware pathfinding and architecture optimization for chiplet-based AI systems Chetan Choppali Sudarshan, Jiajun Hu, Aman Arora, Vidya A. Chhabria Abstract The exponential growth of AI has created unprecedented demand for computational resources, pushing chip designs to the limit while simultaneously escalating the environmental footprint of computing. As the industry transitions toward heterogeneous integration (HI) to address the yield and cost challenges of monolithic scaling, minimizing the carbon cost of these complex HI systems becomes critical. To fully exploit HI, a co-design approach spanning application, architecture, chip, and packaging is essential. However, this creates a vast design space with competing objectives, specifically the trade-offs between performance, cost, and carbon footprint (CFP) for sustainability. CarbonPATH is an early-stage pathfinding framework designed to address this multi-objective challenge. It identifies optimized HI systems by co-designing workload mapping, architectural parameters, and packaging technologies, while treating sustainability as a first-class design constraint. The framework accounts for a wide range of factors, including compute and memory sizes, chiplet technology nodes, communication protocols, integration style (2D, 2.5D, 3D), operational CFP, embodied CFP, and interconnect type. Using simulated annealing, CarbonPATH explores this high-dimensional space to identify solutions that balance traditional metrics against environmental impact. By capturing interactions across applications, architectures, chiplets, and packaging, CarbonPATH uncovers system-level solutions that traditional methods often miss due to restrictive assumptions or limited scope. I Introduction Motivation and problem. As AI systems scale to unprecedented sizes, designing hardware that balances performance with environmental sustainability is becoming a fundamental challenge. These models require executing trillions of operations and moving terabytes of data, which far exceeds the capability of a single accelerator. The most performance-efficient solution would be a monolithic chip that accommodates all model parameters and intermediate activations on a single die [ 35 ] . However, the slowdown of Moore’s law and Dennard scaling, combined with reticle size limits and escalating costs, makes this approach impractical and unsustainable [ 41 , 45 ] . The manufacturing of these massive monolithic dies increases environmental impact, as lower die yields result in a larger manufacturing carbon footprint (CFP) and increased silicon waste [ 41 , 45 ] . Advanced packaging and heterogeneous integration (HI) have therefore emerged as promising alternatives [ 19 ] . By assembling multiple specialized dies within a single package, designers can approximate the performance of a large monolithic accelerator while improving yield, reducing non-recurring engineering costs, reusing IP across product generations [ 13 ] , and lowering CFP [ 45 ] . The design of chiplet-based AI accelerators, however, presents significant challenges. System-level performance, power or energy efficiency, area, cost (PPAC), and CFP depend not only on application and architectural choices, such as resource allocation, workload mapping, on-chip memory sizes, sizing of compute, and dataflow, but also on chip and packaging decisions, including placement, interconnect topology, and the choice between 2.5D and 3D integration technologies. These choices are highly interdependent. For example, modifying resource allocation affects bandwidth requirements, which in turn constrains packaging choices and impacts overall system cost and CFP. Most existing research often addresses architecture and packaging independently [ 35 , 51 , 9 , 27 , 40 , 46 , 26 ] , which leads to suboptimal designs for HI systems. Fig. 1: System-level co-design across application, chip architecture, chiplet configuration, and package choices for chiplet-based accelerators. Further, as major technology firms increasingly commit to strict decarbonization timelines [ 2 , 14 , 29 ] , minimizing total CFP has moved from a corporate initiative to a fundamental engineering constraint. Total CFP comprises embodied emissions from the manufacturing of dies and packaging of the HI system, and operational emissions from active workload execution. While HI offers a pathway to lower embodied carbon via improved yields [ 45 ] , it must not come at the cost of system performance. Therefore, achieving true sustainability requires a holistic approach where carbon is treated as a primary optimization objective, co-optimized with traditional PPAC during the early stage architectural definition stages. The complexity of HI systems requires co-design across the compute stack, including application, architecture (compute and memory), chip, and package space. This system-level co-design is highlighted in Fig. 1 , which illustrates the variety of design considerations that must be made at each layer of the stack. These considerations are also dependent on other decisions, blurring the boundaries between the layers of the stack. The entire design space is combinatorially large, making exhaustive search infeasible. TABLE I: Comparison of prior modeling and optimization frameworks. Framework Chiplets of different sizes Cycle-accurate performance D2D Protocol- aware Dataflow- aware Workload mapping Chiplet floorplanning DSE Dollar cost Embodied CFP Operational CFP HISIM [ 52 ] ✗ ✗ ✗ ✓ ✗ ✗ ✓ ✗ ✗ ✗ ACT [ 18 ] ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ECO-CHIP [ 45 ] ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ CATCH [ 16 ] ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ChipletGym [ 30 ] ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✗ CarbonPATH [ 44 ] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Prior work. A summary of prior work is provided in Table I . The table compares frameworks across key HI system co-design capabilities, spanning modeling assumptions, system awareness, and optimization scope. Specifically, we distinguish whether a framework supports chiplets of different sizes and whether it uses cycle-accurate performance models versus higher-level/analytical modeling. We also indicate whether the framework is aware of die-to-die (D2D) protocol constraints and dataflow characteristics. We highlight support for workload-to-chiplet mapping (exploring multiple mapping strategies) and chiplet floorplanning (optimizing chiplet placement and the implied D2D connectivity). We report whether each framework performs design-space exploration (DSE) and optimization, and whether it models economic and sustainability metrics, including dollar cost and embodied/operational carbon footprint (CFP). HISIM [ 52 ] presents an analytical modeling framework for the performance analysis of heterogeneous 2.5D/3D integrated AI systems, enabling efficient DSE. [ 18 ] leverages publicly available data to create data-driven models for embodied CFP, whereas ECO-CHIP [ 45 ] has modeled the overall CFP for chiplet-based systems, including both embodied and operational CFP (does only CFP analysis, no optimization). CATCH [ 16 ] provides cost models for multi-chiplet based designs, accounting for packaging, testing, and non-recurring engineering costs. ChipletGym [ 30 ] addresses the challenge of the vast design space by exploring it via the use of reinforcement learning (RL) and simulated annealing (SA). However, it falls short in several key aspects: (1) inaccurate modeling of PPAC due to unrealistic simplifying assumptions, (2) not including CFP as an optimization objective, (3) missing specific design considerations that lead to a narrower design space, (4) does not analyze the impact of diverse workload mapping styles, and (5) it fails to optimize across the entire stack. All the above aspects impact the design of HI systems. CarbonPATH contributions. CarbonPATH introduces a carbon-aware architecture exploration framework for chiplet-based AI systems that jointly considers workload mapping, chiplet architecture, and advanced packaging technologies. Unlike prior approaches that focus on performance or cost alone, CarbonPATH treats carbon footprint—both embodied and operational—as a first-class design objective and explicitly models its interaction with architectural and packaging decisions. 1. Carbon-aware cross-layer architecture