Artificial Intelligence Experts Construct Lego-Style Microchip
In a groundbreaking development, MIT engineers have unveiled a Lego-like design for a stackable, reconfigurable artificial intelligence (AI) chip. This innovative research, partly funded by the Ministry of Trade, Industry, and Energy (MOTIE) from South Korea, the Korea Institute of Science and Technology (KIST), and the Samsung Global Research Outreach Program, promises to revolutionise edge computing and wearable devices.
The design, reminiscent of Lego blocks, consists of standardised building blocks or tiles that can be combined in various ways to create tailored AI hardware. This modularity allows for customisation for specific AI tasks and adaptability to changing needs. The stacked modules communicate efficiently to perform neural network computations.
The compactness, flexibility, and low power consumption of this chip make it ideal for integration into resource-constrained environments like edge nodes, mobile devices, and wearables. This integration enables on-device AI inference and sensor integration closer to the source of data, reducing latency and dependence on cloud connectivity.
The potential applications are vast, ranging from personalised health monitoring, real-time environmental sensing, and various Internet of Things (IoT) applications in wearables and edge systems. Jeehwan Kim, one of the researchers, even suggests the possibility of selling different types of neural networks separately, allowing customers to choose what they want and add to an existing chip like a Lego.
The team, including MIT postdoc Jihoon Kang and several other authors and collaborators from various institutions, has published their results in Nature Electronics. The design currently carries out basic image-recognition tasks, with each array in the chip trained to process and classify signals directly on the chip, without the need for external software or an Internet connection.
The design consists of alternating layers of sensing and processing elements, with light-emitting diodes (LED) for communication between layers. The team used artificial synapses, arrays of memory resistors, to function as a physical neural network in the chip. The optical communication system consists of paired photodetectors and LEDs, each patterned with tiny pixels.
The team fabricated a single chip with a computing core measuring about 4 square millimeters, stacked with three image recognition "blocks." Each block comprises an image sensor, optical communication layer, and artificial synapse array for classifying one of three letters, M, I, or T. The chip can be reconfigured by swapping out or stacking layers to add new sensors or updated processors.
One of the researchers, Choi, previously developed a "smart" skin for monitoring vital signs. The team plans to add more sensing and processing capabilities to the chip, with potential applications in smartphones, healthcare monitors, and modular electronics.
Despite its initial limitations in distinguishing between blurry images, the team remedied this issue by swapping out the chip's processing layer for a better "denoising" processor. Each neural network array in the chip produces a larger electrical current in response to clear images of the letters it is trained to recognise.
In summary, MIT's modular AI chip design offers a promising solution for creating adaptable AI hardware suitable for compact, power-sensitive edge and wearable devices. The Lego-like approach allows for customisation, efficient AI hardware, and potential applications in various industries, from healthcare to environmental monitoring and IoT devices.
- The MIT engineers' Lego-like design for a stackable, reconfigurable artificial intelligence (AI) chip is funded by MOTIE, KIST, and the Samsung Global Research Outreach Program.
- The design, consisting of standardized building blocks, is reminiscent of Lego blocks and allows for customization for specific AI tasks.
- The modularity of the design allows for adaptability to changing needs, with the stacked modules communicating efficiently to perform neural network computations.
- The compactness, flexibility, and low power consumption of this chip make it ideal for integration into resource-constrained environments like edge nodes, mobile devices, and wearables.
- The potential applications of the chip are vast, ranging from personalized health monitoring, real-time environmental sensing, to various Internet of Things (IoT) applications in wearables and edge systems.
- The team has published their results in Nature Electronics and is planning to add more sensing and processing capabilities to the chip, with potential applications in smartphones, healthcare monitors, and modular electronics.
