AC-Copilot: Application-Algorithm-Hardware Co-design Workflow and Toolchain

Background

The emergence of large-scale AI models has created growing demand for high-efficiency accelerator chips, driven by their computational power and energy-saving capabilities. However, conventional chip design cycles are often slow and complex, requiring deep expertise in AI algorithms and domain-specific knowledge to achieve optimal deployment. AC-Copilot addresses these challenges through a unified flow that integrates applications, algorithms, and hardware design. By combining modules for neural architecture search, model compression, hardware simulation, and software generation, it enables fast, flexible, and cost-effective development of AI accelerators.

Innovation

AC-Copilot brings deep innovation to AI accelerator design through its specialized high-efficiency data path and engine architecture templates tailored for deep learning workloads. Its coverage spans the full development stack—from popular DL frameworks like PyTorch and ONNX to detailed hardware implementation—offering seamless integration across diverse application scenarios. It leverages large language models (LLMs) to enhance performance estimation accuracy. This intelligent synergy between AI and hardware co-design makes AC-Copilot a powerful, end-to-end solution for building next-generation AI chips.  

Competitiveness

The AC-Transformer, developed with AC-Copilot and fabricated on a 28nm process at 500MHz, demonstrates exceptional results when tested on the EfficientVit model. Compared to a commercial 8nm AI chip running at 625MHz, it achieves 16.3× higher compute efficiency and 7.2× better energy efficiency. Additionally, a planned 12nm prototype designed through AC-Copilot shows an even greater 19.8× improvement in energy efficiency on the same model.

These comparisons underline AC-Copilot’s strength in optimizing chip performance tailored to modern AI algorithms.  

Applications

AC-Copilot delivers high-performance AI accelerator designs with integrated SDKs and runtime—significantly reducing development and deployment overhead. Its broad adaptability suits various industries: Semiconductor Companies: Streamline custom AI chip creation with minimal software toolchain effort. EDA Vendors: Plug in to enhance automated architecture design and optimization flows. AI Solution Providers: Evaluate the benefits of dedicated hardware to boost solution performance.