Title: Uncertainty-Aware End-to-End Co-Design of Neural Network Processors: From Training and Mapping to Fabrication

Abstract:

The development of neural network processors constitutes a comprehensive end-to-end co-design challenge. In this context, the inference workload is dictated by the network architecture and training budget; chip area, latency, and energy consumption are determined by hardware mapping choices; and these technical attributes ultimately influence fabrication yield and manufacturing expenses. Currently, these decisions are typically isolated into distinct phases, and prevailing co-design methods are often rigidly tied to particular algorithms. This rigidity hinders the ability to enhance individual components without necessitating a complete overhaul of the entire pipeline.

To address these limitations, this study introduces a unified framework rooted in monotone co-design theory. This framework integrates four interoperable design blocks that cover network training, chip mapping, wafer-level fabrication, and compute resource allocation. By exposing only a functionality-resource interface to the broader system, each block can be optimized independently without requiring structural modifications to other components. A primary innovation of this approach is its handling of uncertainty. Instead of reducing stochastic outcomes to simple point estimates, the framework treats "Confidence"—defined as the inverse of success probability—as an explicit, optimizable resource comparable to cost, time, and power.

The efficacy of this methodology is demonstrated through three case studies. The first validates the framework’s ability to identify Pareto-optimal implementations across diverse application scenarios. The second confirms that Confidence serves as a continuously adjustable design parameter rather than merely a retrospective diagnostic tool. Finally, the third case illustrates that enhancing the implementation set of any single block automatically improves the global Pareto front, achieving this result without altering the underlying co-design architecture.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC