New Energy-Efficient AI Chip is Inspired by the Brain
This post is also available in: 
עברית (Hebrew)
Artificial intelligence is powered by great amounts of electricity, and in order to run complicated applications like behavior monitoring, facial recognition software, or live object tracking in real time, there needs to be a computing system with faster and more accurate inferences. For this to happen, a large AI model must work closely with the source of data.
The problem of moving large amounts of data between computer and memory is not new by any means, and in order to solve it, researchers have developed an architectural innovation that combines brain-inspired computing and semiconductor technology. “NorthPole” consumes less energy and processes data efficiently as it intertwines computing with memory on a single chip.
According to Interesting Engineering, there were three factors motivating the researchers when creating the NorthPole- neural inference (inspired by the brain and can be implemented with less complicated constructs), Moore’s Law (according to which the number of transistors on a chip doubles every 10 years) and preventing artificial intelligence from becoming an unsustainable technology.
The researchers explained: “Devices that monitor and track in real-time, for example, must communicate with powerful centralized AI models and then bring information back to the edge. This route is fraught with bandwidth limitations, latency problems, and potential disruption to networks.”
The new NorthPole has 22 billion transistors in an 800-millimeter square area, with 256 cores, and blends brain-inspired computing and semiconductor technology, with researchers stating it achieves higher performance, energy efficiency, and area efficiency compared to other comparable architectures, including those that use more advanced technology processes.
In comparison to a graphics processing unit (GPU), NorthPole achieved 25 times higher energy metric of frames per second (FPS) per watt, five times higher space metric of frames per second/transistor, and 22 times less time metric of delay/latency.
The researchers concluded: “Most ambitiously, it is now possible to envision a large network of NorthPole chips, each representing a region in the brain, interconnected via brain-like long-distance pathways and communicating via tensors.”
