Sandia National Labs Deploys SpiNNaker2: A Breakthrough in Energy-Efficient Neuromorphic Computing

Neuromorphic computing, a paradigm that mimics the human brain's structure and function, has been receiving considerable attention from major players like Intel, IBM, and Google, as well as smaller startups. The primary goal is to develop hardware that can perform complex tasks with high efficiency, particularly in areas like AI, edge computing, and IoT, where power consumption is a significant concern. One company at the forefront of this technological shift is SpiNNcloud, a four-year-old German firm that emerged from the Dresden University of Technology. SpiNNcloud is driven by the SpiNNaker2 chip, designed by Steve Furber, a pioneer in the development of the Arm microprocessor. This week, SpiNNcloud announced that Sandia National Laboratories, a leading research institution, has deployed the SpiNNaker2 system. Sandia is well-versed in neuromorphic computing, having already integrated Intel’s Loihi 2 neuromorphic processor into its Hala Point system last year. The deployment of SpiNNaker2 further solidifies their commitment to exploring energy-efficient AI applications. The SpiNNaker2 system at Sandia consists of 24 boards, each housing 48 chips. These chips are interconnected in a toroidal topology, creating a highly parallel architecture capable of simulating about 175 million neurons. Each microchip contains 152 Arm-based, low-power processing elements, making the system extremely resource-efficient compared to traditional GPUs. According to SpiNNcloud, SpiNNaker2 is 18 times more efficient than current AI inference GPUs, with its successor, SpiNNext, expected to be 78 times more efficient. Hector Gonzalez, co-founder and CEO of SpiNNcloud, highlighted several advantages of SpiNNaker2. The system's globally asynchronous, locally synchronous design allows for fine-grained control over each of the 175,000 cores, enabling users to isolate and control individual paths more effectively than in GPU-based systems. Additionally, SpiNNaker2’s flexibility allows it to support both event-based neuromorphic computing and mainstream deep neural networks (DNNs). This versatility is crucial for scaling neural symbolic models, which combine symbolic reasoning with neural layers. Sandia scientists will use the SpiNNaker2 system to explore a range of applications. One key application is in drug discovery, where small multilayer perceptrons (MLPs) can be deployed at scale to match molecular patterns and patient profiles. This approach leverages parallel processing to enhance the efficiency and speed of drug discovery processes. Other potential uses include solving QUBO (Quadratic Unconstrained Binary Optimization) problems, which are common in logistics and complex mathematical simulations. By deploying workers at scale using random algorithms, the system can explore multiple solutions simultaneously, making it ideal for optimizing logistical operations and handling complex mathematical challenges. Gonzalez also emphasized the importance of dynamic sparsity in AI computing. Recent advancements in machine learning have moved the industry from dense modeling to extreme dynamic sparsity, where only relevant neural pathways are activated based on input. This approach significantly reduces computational work and energy consumption. However, standard hardware like GPUs is not well-suited for this level of fine-grained isolation.SpiNNaker2’s ability to execute only the required parts of a network—known as a mixture of experts—makes it an excellent fit for generative AI algorithms and dynamic sparsity. Industry insiders view the deployment of SpiNNaker2 at Sandia as a promising step toward mainstreaming neuromorphic computing. The system’s power efficiency and flexible design could address the energy-intensive nature of current AI and DNN applications, making it a valuable asset in the pursuit of sustainable and scalable computing solutions. SpiNNcloud, with its roots in academia and a focus on innovation, is positioned to play a significant role in shaping the future of neuromorphic computing.