Orthrus: Dual-Loop Automated Framework for System-Technology Co-Optimization

Yi Ren, Baokang Peng, Chenhao Xue, Kairong Guo, Yukun Wang, Guoyao Cheng, Yibo Lin, Lining Zhang, Guangyu Sun

Abstract

As Moore's Law slows, the traditional separation between chip architecture and semiconductor manufacturing limits further performance gains. This paper introduces Orthrus, an automated framework utilizing a dual-loop optimization strategy to bridge the gap between system-level design and technology-level parameters. By integrating system-level architectural exploration with technology-level transistor and cell tuning, Orthrus achieves significant power and performance improvements, demonstrating up to 61.4% power savings or a 12.5% performance increase in 7nm predictive designs.

1. Introduction

For decades, the semiconductor industry has relied on a decoupled "Fabless-Foundry" model where designers focus on logic and architecture while foundries optimize underlying transistor physics. This separation is mediated by the Process Design Kit (PDK). While efficient, this isolation prevents designers from exploiting technology-level tuning to solve specific system-level bottlenecks.

System-Technology Co-Optimization (STCO) aims to break these silos. However, STCO is hindered by information gaps—technology developers lack insight into which circuits drive system performance, and architects cannot easily navigate physical manufacturing parameters. Furthermore, the computational cost of evaluating the entire stack from RTL to physical layout precludes exhaustive search. Orthrus addresses these hurdles through an automated, dual-loop approach.

2. The Orthrus Framework

Orthrus addresses complexity through a "dual-loop" architecture, named after the two-headed dog of Greek mythology. This structure allows for concurrent optimization of system and technology parameters while maintaining a feedback bridge between them.

The Orthrus dual-loop optimization framework, illustrating the System Loop, Technology Loop, and the Inter-Loop Direction mechanism.

2.1 The System Loop

The System Loop focuses on high-level design choices such as clock frequency, synthesis effort, and placement density. It utilizes Multi-Objective Bayesian Optimization (MOBO) to efficiently explore the design space and identify the Pareto frontier—the set of optimal trade-offs between Power, Performance, and Area (PPA).

2.2 The Technology Loop

The Technology Loop optimizes the fundamental building blocks of the library. It adjusts transistor work functions, fin dimensions, and cell layouts. To overcome the high latency of traditional Electronic Design Automation (EDA) tools, Orthrus employs a Neural Network (NN) surrogate model. This model acts as a high-speed proxy for transistor-level simulations, allowing an Enhanced Differential Evolution (DE) algorithm to evaluate thousands of configurations in seconds.

3. Inter-Loop Communication and Cell Fusion

The core innovation of Orthrus is the "Inter-Loop Bridge," which translates system-level requirements into actionable technology-level targets. This bridge performs three critical functions:

  1. Cell Contribution Analysis: It identifies which standard cells (e.g., XOR, Adders) reside on the critical timing paths of the specific system.
  2. Subcircuit Mining: It detects frequent netlist patterns to create custom "fused" multi-row standard cells, reducing wire length and parasitics.
  3. Directional Guidance: It calculates the "normal vector" of the system's Pareto frontier, directing the technology loop toward specific PPA goals (e.g., prioritizing speed over power).
Comparison of standard cell layouts (a) and optimized fused multi-row configurations (b, c) generated to reduce parasitics and wire length.

By identifying that Full Adders (FAs) and Half Adders (HAs) accounted for over 50% of delay in their test system, the authors used the StdGen tool to "fuse" these into custom layouts, bypassing the rigid constraints of standard single-row cell heights.

4. Experimental Results

The framework was validated using the ASAP7 7nm Predictive PDK and a target system consisting of an 8×8 systolic array of Multiply-Accumulate (MAC) units. Orthrus demonstrated a 33.2% improvement in Hypervolume, significantly expanding the achievable design space compared to traditional baseline optimization. Key PPA gains include:

  • Iso-Power Performance: A 12.5% reduction in delay at the same power budget.
  • Iso-Delay Efficiency: A 61.4% reduction in power consumption at the same target speed.

Validation using cosine similarity confirmed that the optimization path taken by the technology loop was highly aligned with the "ideal" direction suggested by the system-level analysis, proving the effectiveness of the inter-loop bridge.

5. Conclusion and Future Work

Orthrus represents a significant step toward automated STCO. By breaking the traditional silos between design and manufacturing, it enables substantial PPA gains unreachable through isolated optimization. Future research will explore the application of Orthrus to more irregular logic structures, such as general-purpose CPUs, and extend the framework to "Beyond-CMOS" technologies like Gate-All-Around FETs (GAAFETs).

Summary updated

BibTeX
@inproceedings{ren2025orthrus,
  author = {Yi Ren and Baokang Peng and Chenhao Xue and Kairong Guo and Yukun Wang and Guoyao Cheng and Yibo Lin and Lining Zhang and Guangyu Sun},
  title = {{Orthrus: Dual-Loop Automated Framework for System-Technology Co-Optimization}},
  booktitle = {ICCAD},
  year = {2025},
  doi = {10.1109/iccad66269.2025.11240673},
}
Figure