IEEE ICDCS 2026

There has been an exponential growth of cyber-physical applications that rely on diverse types of embedded end-systems and devices, such as smart phones/watches/glasses, home appliances, consumer and industrial electronics, smart sensors and actuators. These applications require diverse types of Quality-of-Service (QoS) including timeliness, dependability, security and privacy, from end-systems/devices which are usually networked together via heterogeneous networking technologies and protocols.

We now know how to guarantee timeliness and, to a lesser extent, how to provide fault-tolerance, on both end-systems and their interconnection networks. However, how to secure them is far less known, despite the growing importance of protecting information stored in the end systems/devices and exchanged over their interconnection networks. Moreover, timeliness, fault-tolerance, security and privacy must be supported simultaneously, often with a tight resource budget such as memory, computation and communication bandwidth, and battery power. This talk will cover issues and approaches to the problems of securing networked embedded systems.

If time allows, I will discuss our work-in-progress on context-aware autonomous vehicles.

Parallel computing is entering a new era, driven by the rapid rise of AI, and undergoing a significant transformation. Previously associated primarily with supercomputers and scientific workloads, it now serves as the foundation of contemporary AI infrastructure. Large-scale training, real-time inference, agentic systems, simulation, robotics, and physical AI require extensive parallelism across accelerators, memory, networks, storage, software stacks, and energy infrastructure.

This talk explores the future direction of parallel computing systems in the context of AI. The next generation of systems will be defined by comprehensive full-stack co-design, rather than incremental improvements in processor speed or cluster size. Industry initiatives such as NVIDIA's AI factories and Google's AI Hypercomputer exemplify this transition. The talk will address major challenges including heterogeneous accelerators, communication bottlenecks, scheduling, reliability, observability, data locality, energy efficiency, sustainability, and openness.

The deployment of Large Language Models (LLMs) on devices faces significant challenges due to their extensive memory requirements. This talk introduces a series of work towards memory-efficient LLM from the perspective of model size as well as from the perspective of KV cache.

Double Compression combines model compression (quantization and pruning) with lossless data compression, achieving a 2.2x compression ratio while maintaining model accuracy within a 1% drop. FlexInfer leverages advanced system techniques such as prefetching and memory locking to maximize memory efficiency, achieving up to 12.5x faster inference under memory constraints compared to traditional methods. Together, these solutions enable memory-efficient LLM deployment on edge devices. The talk will also present the recent project ClawMobile, which runs efficient agents on mobile devices.