A new chip architecture developed by IBM Research could significantly improve the efficiency of AI computing. The prototype chip, called NorthPole, was created by IBM researcher Dharmendra Modha and his team after nearly two decades of work on brain-inspired computing.
According to the research published in Science, NorthPole achieves much higher performance and efficiency compared to traditional chips like GPUs and CPUs for AI inferencing tasks. In tests with image recognition and object detection models like ResNet-50 and YOLOv4, NorthPole was found to be up to 25 times more energy efficient and 4,000 times faster than previous neuromorphic chips like TrueNorth.
NorthPole’s speed and efficiency stems from its unique design that integrates memory with computation, eliminating the von Neumann bottleneck that hinders most chips today. This allows NorthPole to perform operations very quickly while using minimal energy. The current NorthPole prototype contains 22 billion transistors and 256 cores for parallel processing.
While NorthPole excels at AI inferencing, its memory capacity is limited to what can fit on the chip itself. However, multiple NorthPole chips can work together in a scale-out configuration for larger models. The team is exploring mapping large language models across these chip clusters.
Possible applications for NorthPole include computer vision tasks like image recognition, video analysis, speech processing, and natural language processing. It could enable more powerful AI capabilities in edge devices like autonomous vehicles, robots, and security systems according to Modha.
The NorthPole project represents a breakthrough in neuromorphic chip design, combining insights from neuroscience with advanced chip engineering. With further improvements to bring NorthPole to smaller process nodes, its efficiency advantages over traditional hardware could greatly accelerate AI adoption across many industries. IBM Research plans to continue optimizing NorthPole’s architecture and demonstrating new applications.
