MIT Engineers Build Stackable AI Chip to Cut E-Waste

MIT researchers have developed a stackable artificial intelligence chip that can be reconfigured for different tasks, potentially extending hardware lifespans and reducing electronic waste in the burgeoning Internet of Things era.

Optical Links Replace Physical Wires

The chip, published in Nature Electronics in 2022, uses light instead of traditional metal wiring to communicate between layers. LEDs and photodetectors transmit data optically, allowing individual layers to be swapped or upgraded without replacing the entire device.

The architecture comprises alternating layers of sensing and processing elements. Each layer contains memristor arrays that function as artificial synapses, mimicking neural network behavior in hardware rather than software.

Addressing IoT Demands

"As we enter the era of the internet of things based on sensor networks, demand for multifunctioning edge-computing devices will expand dramatically," said Jeehwan Kim, MIT associate professor of mechanical engineering.

Current AI hardware faces a fundamental mismatch: machine learning models evolve rapidly, but physical chips remain static. Devices become obsolete not because they fail, but because they cannot adapt to new algorithms or workloads. This drives premature replacement cycles and contributes to growing e-waste streams.

Sustainability Through Reconfigurability

The stackable design directly targets this problem. When AI requirements change, users could add or swap layers rather than discard entire devices. A sensor node deployed for temperature monitoring could be reconfigured for vibration analysis by exchanging processing layers while retaining the same sensing hardware.

This approach matters most at the edge, where billions of IoT devices operate in remote or embedded locations. Physical replacement becomes costly and environmentally burdensome at scale. Reconfigurable hardware offers a path to keep devices current through their full mechanical lifespan.

Technical Foundation

The optical communication scheme eliminates bandwidth bottlenecks that plague traditional stacked chips. Physical interconnects create heat and limit how many layers can be practically stacked. Light-based data transfer generates less heat and supports denser layer configurations.

Memristor arrays provide the reconfigurable computing element. These devices remember their resistance state even without power, enabling them to store both data and processing logic. The same physical hardware can be programmed to execute different neural network architectures as needs evolve.

What Comes Next

The Nature Electronics publication establishes the feasibility of optically-connected stackable AI hardware. Commercial deployment will require manufacturing scale-up and integration with existing IoT ecosystems. If successful, the technology could reshape how edge devices are designed, purchased, and maintained—treating chips as upgradeable components rather than disposable commodities.