Problem

The rapid advancements in AI models and hardware are currently fragmented, leading to significant inefficiencies. AI algorithms are often designed without considering hardware limitations, while hardware is optimized for past workloads. This disconnect results in escalating energy consumption for large AI models, which is environmentally and economically unsustainable. The lack of a cohesive long-term vision for AI+HW co-development hinders the creation of holistic, adaptive AI systems capable of operating efficiently across various environments, from cloud to edge.

Method

This paper proposes a 10-year roadmap for AI+HW co-design, focusing on three abstraction layers: hardware technologies, algorithms and paradigms, and applications and societal impact. Key approaches include shifting to memory-centric and 3D-integrated architectures to reduce data movement bottlenecks. It advocates for AI-in-the-loop design workflows, where AI assists in hardware design and hardware becomes AI-adaptive. The method emphasizes continuous cross-layer optimization, developing low-complexity yet high-quality AI models, and re-evaluating success metrics from raw computational power to "intelligence per joule."

Results

The vision outlines ambitious goals for the next decade, including a 1000x improvement in AI training and inference efficiency. This will be achieved through a combination of algorithmic advancements (10x), silicon utilization and advances (20x), and system-level efficiency (5x). It foresees a shift towards energy-aware, self-optimizing systems that seamlessly integrate across cloud, edge, and physical environments. The paper predicts a fundamental distribution shift where smaller, specialized models (<20B parameters) will dominate edge and on-device physical AI applications, complemented by large "teacher" models in the cloud. Success will also encompass democratized access to AI infrastructure and human-centric design principles.

Implications

Realizing this vision requires profound collaboration among academia, industry, and government, fostering coordinated national initiatives, shared infrastructure, and robust workforce development. The AI+HW co-design will unlock new applications in scientific discovery, healthcare, climate modeling, and autonomous systems. It is crucial to address challenges such as the impending power crisis and talent gaps to ensure U.S. competitiveness. By aligning technological progress with sustainability and societal benefit, this integrated approach aims to create an AI ecosystem that is not only powerful and efficient but also reliable, secure, and socially responsible.