Special Sessions

Design and Optimization for AI/ML Acceleration on Resource-constrained Systems

  • Time: Sept. 30, 13:30~15:00
  • Organizers:
    Yuan-Hao Chang (Academia Sinica)
    Jalil Boukhobza (ENSTA Bretagne)

The rapid advancement of AI/ML (including LLM) and edge computing has placed unprecedented demands on computation, memory, and storage of resource-constrained systems. As AI/ML models scale, the ability to efficiently manage computing resources, utilize memory and storage, and reduce energy consumption has become critical. This special session brings together experts in the fields to explore novel architectures and algorithms, memory/storage-centric optimizations, and hardware-aware codesign approaches, addressing the key bottlenecks in AI/ML, for both the learning and inference task. Our aim is to provide attendees with insights into cutting-edge research and practical solutions for optimizing memory and storage in AI/ML for edge computing applications. By addressing both hardware and software perspectives, this session will foster discussions on how to build more efficient, scalable, and cost-effective AI/ML systems.

  1. Hardware-aware DNN Architecture and Mapping Co-optimization for Efficient Inference on Resource-constrained Heterogeneous Systems: Daniele Jahier Pagliari, Alessio Burrello (Politecnico di Torino)
  2. Scaling RAG on Resource-constrained Systems: Advanced Memory, Storage, and Energy-efficient Designs for Next-Gen AI: Chun-Feng Wu (National Yang Ming Chiao Tung University)
  3. Towards Cost-effective and High-performance Large-Scale Graph Processing on Resource-constrained Systems: Ming-Chang Yang (The Chinese University of Hong Kong)
  4. Learning on the Edge: Unlocking the Storage Bottleneck with a Divide and Conquer Approach for Resource-constrained Edge Systems: Jalil Boukhobza (ENSTA Bretagne)

Predictable Timing Behavior in Distributed Cyber-Physical Systems

  • Time: Sept. 30, 15:30 – 17:00
  • Organizer: Jian-Jia Chen (TU Dortmund University)

Ensuring predictable and deterministic behavior in distributed cyber-physical systems (CPS) is essential for guaranteeing safety, reliability, and real-time behavior. However, achieving this predictability is challenging due to network uncertainties, asynchronous execution, and complex timing interactions. This special session brings together experts to explore in four presentations how this uncertainty can be addressed and how to introduce additional determinism into the system to achieve predictable timing behavior in distributed CPS.

We begin by exploring the cornerstones of timing analysis techniques for providing end-to-end latency guarantees for distributed systems (Chen and Günzel). Next, we discuss design strategies for meeting timing constraints, focusing on how system parameters influence cause-effect chains and how these parameters can be tuned to ensure predictable behavior in industrial automation settings (Dasari and Becker). We then turn to approaches for achieving more predictable system behavior. To that end, we examine deterministic semantic models for distributed systems that enable the design of robust and fault-tolerant systems (Lee). Finally, we discuss how solving constraints for scheduling dataflow synchronous programs can be used to enforce strict timing guarantees and improve predictability (Bourke).

  1. Cornerstones in Analytical End-to-End Timing Analysis: Jian-Jia Chen, Mario Günzel (TU Dortmund University)
  2. Design Strategies to Meet End-to-End Timing Requirements of Cause-Effect Chains: Dakshina Dasari (Bosch), Matthias Becker (KTH Royal Institute of Technology)
  3. Why Determinism Matters in Distributed CPS: Edward A. Lee (UC Berkeley)
  4. Solving Constraints to Schedule Dataflow Synchronous Programs: Timothy Bourke (Inria Paris)

Intermittent TinyML: Powering Sustainable Deep Intelligence Without Batteries

  • Time: Sept. 30, 15:30 – 17:00
  • Organizer: Pi-Cheng Hsiu (Academia Sinica)

Energy harvesting enables low-cost, tiny devices to be powered by ambient energy sources without the need for batteries. However, ambient energy is inherently weak and unstable, causing frequent power failures in battery-less devices and forcing them to operate intermittently. This makes executing resource and energy-intensive applications like deep neural network (DNN) inference particularly difficult, requiring novel strategies for safe and efficient execution. This proposed special session will explore crucial challenges in realizing intermittent TinyML from diverse research perspectives, including runtime software, design time tools, multi-device net working, and specialized applications, while also examining their sustainability implications.

  1. Efficient and Sustainable Deep Inference on Intermittent Battery-less Tiny Devices: Hashan Roshantha Mendis (Academia Sinica)
  2. Algorithms and Architectures for Intermittent Inference on Battery-less Sensors: Kasim Sinan Yildirim (University of Trento)
  3. Methods and Tools for Battery-less Intermittent Networks: Marco Zimmerling (TU Darmstadt)
  4. Building Up to Intermittent Inference in Space: Luca Mottola (Politecnico di Milano)

Hardware-Software Co-Design for Machine Learning Systems Made Open-Source

  • Time: Sept. 30, 15:30 – 17:00
  • Organizers:
    Mehdi Tahoori (Karlsruhe Institute of Technology)
    Jörg Henkel (Karlsruhe Institute of Technology)
    Jürgen Teich (Friedrich-Alexander-University erlangen-nürnberg)

Chip technologies are crucial for the digital transformation of industry and society. Machine Learning (ML) and Artificial Intelligence (AI) are increasingly shaping both daily life and industrial applications, with AI hardware playing a vital role in enabling efficient and scalable ML deployment. However, significant challenges remain in bridging the gap between ML algorithm development and hardware implementation, particularly for edge ML applications where efficiency, power constraints, and adaptability are critical. One of the key bottlenecks in ML hardware development is the lack of seamless integration between ML toolchains and electronic design automation (EDA) tools for hardware synthesis and mapping. Current solutions often require extensive manual optimization and costly proprietary software, limiting accessibility and innovation. Open-source tools can play a transformative role in democratizing ML hardware design, fostering collaboration, and addressing the growing shortage of skilled professionals.

This special session will cover key aspects of hardware-software co-design for ML systems, with a focus on open-source solutions. The session will highlight the critical need for open-source toolchains that connect ML model development with hardware synthesis and optimization. The speakers, representing leading universities across Germany, bring expertise in ML algorithms, hardware design, compiler technologies, and system security. By fostering collaboration between ML and hardware communities, this session aims to accelerate innovation in edge AI and establish a robust open-source ecosystem for ML hardware development.

  1. Accelerator IP Development and Safety Extensions in Open-Source AI Hardware: Vincent Meyer, Mahboobe Sadeghipourrudsari, Mehdi Tahoori, Julian Hoefer, Juergen Becker (Karlsruhe Institute of Technology)
  2. Design Space Exploration of Hardware Architectures and DRAM Interfaces for Optimized AI Systems: Hassan Nassar, Zeynep Demirdag, Heba Khdr, Jörg Henkel, Lukas Steiner, Norbert Wehn (Karlsruhe Institute of Technology; Rheinland-Pfalz Technical University)
  3. Co-Design of AI Applications: ML Compiler and Accelerator Units: Batuhan Sesli, Muhammad Sabih, Frank Hannig, Philipp van Kempen, Johannes Geier, Ulf Schlichtmann, Jürgen Teich (Technical University of Munich; University of Erlangen-Nuremberg)

 

Emerging Scope and Design Challenges for Approximate Computing: Optimizing Accuracy-PPA trade-offs and Beyond

  • Time: Sept. 30, 15:30 – 17:00
  • Organizer: Akash Kumar (Ruhr-Universität Bochum)

In today’s landscape of ever-escalating AI workloads, Approximate Computing (AxC) has emerged as an interesting research topic with the potential for enabling low-cost inference in resource constrained embedded systems. AxC aims to achieve disproportion ate gains in the Power-Performance-Area (PPA) of a system by the deliberate introduction of some form of inaccuracies. Finding the approximate designs that provide the optimal application-specific accuracy-PPA trade-offs still drives most of the related research. AxC can be used across multiple layers to leverage the implicit error tolerance of AI computing. However, this requires adapting the popular AI methods and frameworks to enable the design, testing, and optimization for neural networks with approximate operators. At the same time, modern AxC research is increasingly addressing non-functional objectives. Innovative strategies are being developed to enhance system reliability by reducing reliance on costly redundancy techniques, thereby ensuring robust performance in diverse, resource-constrained environments. Furthermore, the intrinsic error behavior in AxC is opening up new security challenges and opportunities where tailored countermeasures can be designed without significantly impacting resource usage. In addition, emerging application-aware methodologies—illustrated by case studies such as spiking neural networks—demonstrate that incorporating application-specific error characteristics into the design of approximate operators can yield superior performance compared to traditional, one-size-fits-all approaches. This special session will unite perspectives from both academia and industry to explore these multifaceted challenges and opportunities. Discussions will focus on AI-guided design methodologies for approximate operators, strategies for balancing efficiency with reliability and security, and the practical benefits of application-aware design in advancing next generation embedded and high-performance computing systems.

  1. AI-driven Accuracy-PPA Optimization for Approximate Computing: Dr. Siva Satyendra Sahoo (IMEC Leuven)
  2. Balancing Efficiency and Reliability: the Role of Approximate Computing: Bastien Deveautour (University of Nantes)
  3. Security-driven Approximate Computing: Dr. Chongyan Gu (Queen’s University Belfast)
  4. Accuracy-driven Approximate Computing: Dr.-Ing. Salim Ullah (Ruhr University Bochum)