Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic framework that utilizes language models as evolutionary operators to optimize neural network architectures for both task accuracy and energy efficiency across various hardware platforms. By treating hardware as a swappable backend and integrating energy models and non-ideality simulators, UH-NAS facilitates comprehensive system-level comparisons without altering the underlying search algorithm. The framework demonstrates superior performance in discovering diverse and robust architectures on optical MZI hardware compared to conventional methods and existing LLM-to-NAS approaches.
UH-NAS reveals that a hardware-agnostic approach can yield more robust and efficient neural architectures than traditional methods, even in the face of physical non-idealities.
Deploying neural networks on unconventional hardware demands architectures that co-optimize task accuracy and platform-specific constraints such as energy cost, physical non-idealities, and numerical precision. Existing neural architecture search (NAS) methods are typically tailored to a single hardware family, limiting cross-platform comparison and generalization. We introduce Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic, LLM-guided NAS framework that integrates language models as evolutionary operators to co-optimize accuracy and inference energy. By exposing hardware as a swappable backend with per-platform energy models, physical constraints, and non-ideality simulators, UH-NAS enables fair system-level comparisons across various backends without modifying the search algorithm. Tested on optical MZI hardware, UH-NAS discovers more diverse, robust architectures than conventional baselines while outperforming existing LLM-to-NAS approaches. Additional ablations on architecture robustness under non-idealities and the role of system prompts highlight the importance of architecture-hardware co-design for emerging computing platforms.