Artificial Intelligence on Chip Workshop

Dr. Allen Cheng (Light Momentum Technology Corp., Taiwan)

Practical Applications of AI and Customized Chips in Edge Computing Scenarios

This lecture explores real-world deployments of artificial intelligence in edge computing, focusing on how customized semiconductor solutions—from AI accelerators to domain-specific SoCs—enable new levels of performance and energy efficiency.

Assoc. Prof. Andy Yu-Guang Chen (National Central University)

AI-Driven Early Routing, DRC Prediction, and Congestion-Aware Pin Assignment for Faster Physical Design Closure

Modern integrated circuit (IC) design faces mounting challenges from increasing layout complexity, stringent design rules, and ever-tighter schedules. Routing congestion and design rule violations (DRVs) discovered late in the flow often trigger costly, time-consuming iterations.

This talk presents recent advances in applying machine learning early in physical design. Ensemble learning and CNN-based frameworks are introduced for predicting routing demand and detailed-routing-level DRV maps from early placement and global routing data, enabling accurate congestion and violation detection at a fraction of traditional runtime. The talk also introduces COPA, a congestion-oriented pin-assignment method that optimizes feedthrough pin placement to mitigate intra-block routing bottlenecks without increasing wirelength. Together, these data-driven techniques act as lightweight add-ons to existing design flows, reducing design iterations and accelerating closure.

Dr. G. R. Chen (Light Momentum Technology Corp.)

Accelerating the Future: Custom Silicon for AI Workloads

As AI expands across domains—from cloud computing to mobile augmented reality—the demand for customized semiconductor solutions continues to grow. This talk explores architectural foundations and design strategies for AI-specific chip design, emphasizing how neural network characteristics such as sparsity, precision scaling, and dataflow shape hardware implementations.

It covers accelerator paradigms for CNN and transformer workloads and discusses optimization techniques including quantization, pruning, and mixed-precision computation. A case study highlights practical approaches to AI SoC design, focusing on compute efficiency and workload-aware architecture. The talk also surveys trends such as 3D integration, AI-assisted hardware generation, and sustainability-driven design.

Prof. Radim Burget (Brno University of Technology)

Collaborative Problem-Solving Using Large Language Models

This lecture examines the shift from monolithic, opaque AI models toward explainable and scientifically robust AI systems, with a focus on multi-agent large language model (LLM) architectures. It highlights how distributing reasoning, verification, and evidence evaluation across cooperating agents can enable more transparent and reproducible conclusions.

The lecture briefly discusses hardware implications, including increased demands on memory, parallelism, and inter-agent communication, underscoring the tight coupling between explainable multi-agent AI and modern computing architectures.

Ing. Jan Remeš (Mycroft Mind, Ltd., Brno)

Semiconductors at the Grid Edge: Case Studies in PV Forecasting and Smart Meter Intelligence

This session explores practical deployments of embedded AI in the energy sector, showcasing how Mycroft Mind bridges the gap between advanced algorithms and resource-constrained hardware under the IPCEI ME/CT framework.

Through the “SkyEye” case study, the session demonstrates high-precision PV production forecasting using sky imagers and hardware-accelerated ConvLSTM models for real-time edge processing. It also presents a “Consumption Predictor” solution for smart meter intelligence, emphasizing that efficient AI extends beyond LLMs by leveraging incremental learning and optimized mathematical models that operate within strict memory and performance limits of industrial semiconductors.