Executive Summary
Artificial Intelligence has become one of the defining technologies of our time. The recent success of large language models (LLMs) and generative AI was enabled by advances in both algorithms and computing hardware. Over the last decade, graphics processing units (GPUs) have played a central role in transforming AI from a research discipline into a commercially viable technology by enabling the massively parallel computation required to train and operate modern neural networks, from computer vision systems to large language models.
Today, however, AI faces a new challenge. While model capabilities continue to improve, the infrastructure equired to train and deploy these models is becoming increasingly expensive. Energy consumption, cooling requirements, memory bandwidth limitations and hardware costs are emerging as major constraints on the future growth of AI.
As a result, the next phase of AI innovation will not be driven solely by larger and more complex models. It will increasingly depend on novel AI accelerators that reduce data movement, improve energy efficiency, and exploit the physical properties of the underlying hardware.
At the same time, AI is gradually moving beyond cloud-based applications into intelligent physical systems including robots, industrial automation systems, smart sensors and autonomous machines. This development, commonly described as Physical AI, introduces additional requirements such as real-time responsiveness, local decision making and energy-efficient operation.
We believe these two trends are closely related. The same hardware characteristics that improve AI efficiency in data centers are also relevant to AI-enabled IoT, Edge AI, and Physical AI systems, where computation increasingly moves closer to sensors, machines, and actuators. This direction aligns with the AI + HW 2035 roadmap [1].
This paper describes how TECHiFAB TiF® Memristor technology provides a path toward addressing both challenges simultaneously through energy-efficient in-memory computing, reconfigurable states, and scalable hardware architectures.