From autonomous vehicles to collaborative robots, drones, and wearable health devices, intelligent systems are increasingly defined by the tight coupling between AI algorithms and computing hardware. These systems must see, decide, and act in real time, such as detecting pedestrians in traffic, grasping objects in cluttered spaces, or monitoring patient vital signs, while operating under severe constraints of latency, power, and reliability. In such contexts, a few milliseconds or milliwatts can determine whether a decision is safe, sustainable, or even possible. Achieving the right balance among accuracy, speed, and energy efficiency has therefore become central to the design of next-generation intelligent machines.
Inspired by the LEGO assembling principles of decomposing real-world complex entities into bricks, plates, and tiles, we aim to build a guidebook for designing optimal ensemble intelligent systems. The framework conceptualizes AI system construction as a strategic game of modular composition, where algorithms, hardware, and execution strategies are treated as interoperable bricks, each characterized by measurable performance attributes and coupling behaviors. Through a game-theoretic co-design engine, the system learns how these bricks interact under varying workloads, uncovering the latent rules that govern trade-offs among accuracy, latency, and energy across single and multi-model execution. Then, the proposed AI-LEGO profiling pipeline implements this vision by converting fine-grained thread events into an algorithm graph that is mapped onto its corresponding hardware graph, generating detailed execution timelines, energy consumption traces, and per-scenario Brick Profiles. Each Brick Profile captures how algorithmic design, input dimension, and hardware utilization interact to shape performance along three coupled dimensions: accuracy, latency, and energy. Together they form a unified lens to characterize the operational efficiency of modern AI systems under realistic and heterogeneous workloads.
