“`html
Neuromorphic computing is a discipline that leverages concepts from neuroscience to create computing systems that emulate the brain’s functionality and structure. To effectively compete with existing computing technologies, it must undergo significant scaling. A recent review in the journal Nature, authored by 23 researchers, including two from the University of California San Diego, outlines a comprehensive roadmap to achieve this objective. The article provides a fresh and practical insight into how we can match the cognitive capabilities of the human brain while maintaining a similar size and power consumption.
“We don’t expect a universal solution for neuromorphic systems at scale; instead, we foresee various neuromorphic hardware options with unique attributes tailored to specific applications,” note the authors.
Neuromorphic computing applications encompass a wide range, including scientific calculations, artificial intelligence, augmented and virtual reality, wearable technology, smart agriculture, and smart cities. Neuromorphic chips could outperform conventional computers in energy and space efficiency, as well as overall performance, offering significant advantages across sectors like AI, healthcare, and robotics. With AI’s energy consumption anticipated to double by 2026, neuromorphic computing provides an encouraging alternative.
“The relevance of neuromorphic computing is particularly acute today as we witness the unsustainable growth of power-hungry AI systems,” stated Gert Cauwenberghs, a Distinguished Professor in the UC San Diego Shu Chien-Gene Lay Department of Bioengineering and a coauthor of the paper.
Dhireesha Kudithipudi, the Robert F. McDermott Endowed Chair at the University of Texas San Antonio and the lead author of the paper, emphasized the critical moment in neuromorphic computing’s evolution. “We have a significant opportunity to develop new architectures and open frameworks for commercial applications. I firmly believe that enhancing collaboration between industry and academia is essential for guiding the future of this area, as evidenced by our diverse team of coauthors.”
Last year, Cauwenberghs and Kudithipudi obtained a $4 million grant from the National Science Foundation to initiate THOR: The Neuromorphic Commons, an innovative research network designed to provide open access to neuromorphic computing hardware and tools, fostering interdisciplinary and collaborative research.
In 2022, a neuromorphic chip created by a team led by Cauwenberghs demonstrated extreme dynamism and versatility without sacrificing accuracy and efficiency. The NeuRRAM chip executes computations directly in memory and can support a broad range of AI applications, all while consuming significantly less energy than standard general-purpose AI computing platforms. “Our article in Nature discusses potential advancements in neuromorphic AI systems using silicon and new chip technologies to achieve the scale and efficiency reminiscent of self-learning capabilities observed in the mammalian brain,” explained Cauwenberghs.
To scale neuromorphic computing effectively, the authors suggest optimizing several essential characteristics, such as sparsity, which is a fundamental aspect of the human brain. The brain grows by initially creating numerous neural connections (densification) and then selectively removing most of them. This method enhances spatial efficiency while preserving high-fidelity information. If emulated effectively, this capability could lead to neuromorphic systems that are more energy-efficient and compact.
“The capacity for expandable scalability and exceptional efficiency stems from the extensive parallelism and hierarchical structure of neural representation. This involves dense synaptic connectivity within neurosynaptic cores—mirroring the brain’s gray matter—paired with sparse global connectivity for neural communication across cores, reflecting the brain’s white matter. This is made possible by high-bandwidth, reconfigurable interconnects on-chip and across hierarchically structured chips,” stated Cauwenberghs.
“This publication highlights the significant potential for utilizing neuromorphic computing at scale for real-world applications. At the San Diego Supercomputer Center, we are developing new computing architectures for the national user community, and our collaborative efforts are paving the way for introducing a neuromorphic resource to this community,” remarked Amitava Majumdar, director of Data-Enabled Scientific Computing at SDSC on the UC San Diego campus and one of the paper’s coauthors.
“`
