Spike-based Dynamic Computing: Pioneering the Future with Asynchronous Sensing-Computing Neuromorphic Chips
In the pursuit of mimicking the remarkable computational capabilities of the human brain, neuromorphic computing has emerged as a revolutionary paradigm. Neuromorphic hardware, inspired by the non-von-Neumann architecture of the brain, integrates neurons and synapses as computational units. This design allows for highly parallel operations, colocated processing and memory, scalability, and event-driven computation. A groundbreaking development in this field is the introduction of an attention mechanism to Spiking Neural Networks (SNNs) in neuromorphic hardware, marking a significant stride towards incorporating complex neural mechanisms of the human brain into computing hardware.
The essence of neuromorphic intelligence is its suitability for edge computing scenarios—owing to its ability to perceive and process information sparsely. This capability is harnessed through a neuromorphic system known as Speck, which stands as a milestone in the co-design of hardware, algorithms, software, and applications. Speck, a sensing-computing neuromorphic System on Chip (SoC) based on the spike-based sparse computing paradigm, demonstrates the integration of sparse sensing and computing in an innovative manner.
At the heart of Speck lies the Dynamic Vision Sensor (DVS), which simulates biological visual pathways by asynchronously generating spikes in response to changes in scene brightness. This asynchronous spike-based event-driven mechanism is pivotal for achieving high-speed, low-latency, and low-power computing. The physical amalgamation of DVS and the neuromorphic chip within Speck introduces three key designs: the DVS-chip interface system design, the SNN convolution core design, and a fully asynchronous logic design—each playing a critical role in enhancing processing speed and efficiency while minimizing power consumption.
An intriguing aspect of Speck’s design is its dynamic response mechanism, which globally regulates the spike firing of neurons based on stimulus, akin to how the human brain allocates attention. This feature allows Speck to ignore irrelevant information dynamically, reducing the computational load without compromising efficiency. As a medium-scale neuromorphic sensing-computing edge hardware, Speck showcases a remarkable blend of low latency, high dynamic range, and sparse sensing capabilities coupled with an event-driven spiking Convolutional Neural Network (sCNN) processor.
One of the standout features of the Speck sensor is its capability to encode incident light intensity temporally on a logarithmic scale through dynamic vision pixels. This event-based vision system transmits only novel information, significantly reducing data stream size and enhancing energy efficiency. The sensor’s design ensures fully independent operation of each pixel, facilitating high-speed, low-power on-chip data communication.
To adapt the raw Address Event Representation (AER) event stream from the sensor to the requirements of the sCNN, a pre-processing stage is essential. The architectural design of Speck, built in a fully asynchronous fashion, offers unparalleled advantages in sparse sensing and computing. This design choice effectively addresses the challenges of latency and power consumption in edge computing vision applications.
Speck represents a monumental leap in neuromorphic computing, bringing us closer to achieving machine intelligence that mirrors the efficiency and efficacy of the human brain. By efficiently integrating multi-level brain mechanisms, Speck not only advances our understanding of neuromorphic computing but also opens new horizons for its application in real-world scenarios. As this technology continues to evolve, it holds the promise of transforming the landscape of computational neuroscience and machine learning, propelling us towards an era of computing that is as dynamic and intelligent as the brain itself.
