AC-Transformer: High Energy-efficient AI Accelerator Chip

ACCESS
Overview

ACCESS has developed a high energy-efficient AI accelerator chip that supports end-to-end Transformer and CNN AI inference. It can be used in a variety of image and language processing applications.

  • AC-Transformer: High Energy-efficient AI Accelerator Chip
Commercialisation opportunities
Technology licensing
Problem addressed

AC-Transformer features a novel architecture that incorporates a hardware-friendly attention mechanism to improve computation and bandwidth efficiency.

Innovation
  • Implement typical Transformer operators with native hardware acceleration to achieve higher energy efficiency.
  • Two papers were presented at DAC2024, the 61st Design Automation Conference.
  • A paper was accepted by ISSCC2025, "Chip Olympics" IEEE International Solid-State Circuits Conference.
Key impact
  • AC-Transformer achieves a significant breakthrough by linearizing and quantizing the original Transformer model with a 5x reduction in model size for storage and a 45x reduction in computation workload while maintaining the original accuracy and inferencing efficiently on the co-designed accelerator.
  • A next-generation co-design design tool chain supports lightweighting and network segmentation of large Transformer models, and optimizes deployment to one or more accelerator chips.
Award
  • The 49th International Exhibition of Inventions Geneva 2024, Silver Medal
Application
  • Edge Computing for Large AI Models
  • Multi-modal AI systems

Specialized AI chips will be a key factor driving the AI revolution.  To cope with this emerging demand on hardware, the AI Chip Center for Emerging Smart Systems (ACCESS) is putting Hong Kong on the global map of AI chip and hardware design.  ACCESS is a tightly coordinated, multi-disciplinary center for advancing integrated circuit (IC) design technologies to enable novel data-centric computing paradigms supporting a wide range of AI applications.  Putting together world-class experts, the Center is focus on the research theme of designing customised AI chips to realise ubiquitous AI applications used throughout society.

The research agenda in ACCESS is organised into four programmes:

  • Enabling Technology for Emerging Computing Systems addresses memory and data bandwidth problems to alleviate the bottlenecks of AI hardware by exploring integration of silicon-compatible emerging technologies with scaled silicon chips.
  • Architecture and Heterogeneous System Integration focuses on exploring different new architectural and system integration solutions for efficient neuromorphic computing on platforms ranging from cloud to smart Internet of Things.
  • AI-Assisted EDA (Electronic Design Automation) for AI Hardware hopes to develop new design methodologies and design automation tools for AI chips.
  • Hardware-Accelerated AI Applications emphasise selected emerging applications for hardware acceleration, exploring system architecture and new design tools specific to the target applications to achieve breakthroughs in AI hardware in speeds and energy efficiency.

For more inforamtion, please check: https://inno-access.hk/

Enquiry
Related Innovations