Spanning Intelligent Computations across Server to Edge Devices
Although pervasive, almost all computing platforms today are synonymous with the ‘von-Neumann’ computer model. A key characteristic of the von-Neumann model is its clear demarcation of a ‘computing unit,’ from the “storage unit”. which is physically separated from a ‘storage unit.’ This inevitably leads to frequent energy and throughput of intensive data movement between the two units. The resulting ‘Memory-Wall Bottleneck’ renders state-of-the-art computing platforms inefficient for Artificial Intelligence (AI) applications. Therefore, it is not surprising that almost all the payloads for AI computations, such as learning and inference, are performed on a remote server. In turn, this creates a ‘Cognitive-Wall Bottleneck,’ wherein edge devices are deprived of intelligent computations and rely on remote server-like compute resources for critical decision making.
At ASIC Lab, we aim to develop alternate high-throughput and energy-efficient computing paradigms using emerging and existing Silicon technologies, thereby making AI computations accessible to both server and edge devices. In addition to charge-based computing, we’re exploring alternate state variables, including electron spin, photonics, and phononics, for unconventional data-intensive processing. Our approach enables a cross-layer optimization across materials, devices, circuits, architecture, and algorithms for future truly pervasive AI applications.