TACO-HDC: Technology/Algorithm Co-optimization for Hyperdimensional In-Memory Computing towards Ultra-efficient and Reliable Edge AI

• Applicants: Professor Hussam Amrouch, Ph.D.; Professor Mehdi B. Tahoori, Ph.D.
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems ; Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 560224982

Artificial intelligence and machine learning have already made a major impact on our societies and economies. However, the existing trend for AI training and inference is becoming unsustainable, since these tasks are extremely compute and energy demanding, leading to a huge carbon footprint. This proposal delves into the potential of coupling state-of-the-art brain-inspired hyperdimensional computing (HDC) algorithms with innovative computation in-memory (CiM) architectures, realized through emerging technologies, to support sustainable AI, from edge to cloud. HDC, inspired by the human brain, employs high-dimensional vectors and random projections for information representation and processing. This technique boasts unparalleled robustness, error tolerance, and low computational complexity, rendering it an ideal lightweight algorithmic solution for edge-AI applications. Additionally, emerging non-volatile memories (NVMs) are a promising group of technologies offering a low-power alternative to CMOS memory. They consume little to no static power and have a smaller footprint. They can also be used as computational memories in the context of CiM, coupled with analog computing, can further boost the performance and energy-efficiency compared to the traditional computing architectures used for AI acceleration. Nevertheless, NVMs remain in their infancy and are more susceptible to variability and imperfection effects, resulting in significant declines in accuracy. Taking a holistic view, this proposal centers on probing the confluence of HDC algorithms (harnessing their unparalleled error resilience) and trailblazing NVM-based CiM (exploiting their energy efficiency superiority). To realize this, we will develop novel technology/algorithm co-optimization strategies that pave the way for reliable, yet highly energy-efficient and high-performance sustainable AI, from edge devices to cloud computing, undeterred by the unavoidable errors stemming from the underlying emerging technologies and analog computing paradigms.

• DFG Programme: Research Grants